High-order Spatial Interactions Enhanced Lightweight Model for Optical Remote Sensing Image-based Small Ship Detection
Extracted source text · arXiv 2304.03812
%% bare_jrnl.tex
%% V1.4b
%% 2015/08/26
%% by Michael Shell
%% see http://www.michaelshell.org/
%% for current contact information.
%%
%% This is a skeleton file demonstrating the use of IEEEtran.cls
%% (requires IEEEtran.cls version 1.8b or later) with an IEEE
%% journal paper.
%%
%% Support sites:
%% http://www.michaelshell.org/tex/ieeetran/
%% http://www.ctan.org/pkg/ieeetran
%% and
%% http://www.ieee.org/
%%*************************************************************************
%% Legal Notice:
%% This code is offered as-is without any warranty either expressed or
%% implied; without even the implied warranty of MERCHANTABILITY or
%% FITNESS FOR A PARTICULAR PURPOSE!
%% User assumes all risk.
%% In no event shall the IEEE or any contributor to this code be liable for
%% any damages or losses, including, but not limited to, incidental,
%% consequential, or any other damages, resulting from the use or misuse
%% of any information contained here.
%%
%% All comments are the opinions of their respective authors and are not
%% necessarily endorsed by the IEEE.
%%
%% This work is distributed under the LaTeX Project Public License (LPPL)
%% ( http://www.latex-project.org/ ) version 1.3, and may be freely used,
%% distributed and modified. A copy of the LPPL, version 1.3, is included
%% in the base LaTeX documentation of all distributions of LaTeX released
%% 2003/12/01 or later.
%% Retain all contribution notices and credits.
%% ** Modified files should be clearly indicated as such, including **
%% ** renaming them and changing author support contact information. **
%%*************************************************************************
% *** Authors should verify (and, if needed, correct) their LaTeX system ***
% *** with the testflow diagnostic prior to trusting their LaTeX platform ***
% *** with production work. The IEEE's font choices and paper sizes can ***
% *** trigger bugs that do not appear when using other class files. *** ***
% The testflow support page is at:
% http://www.michaelshell.org/tex/testflow/
\documentclass[journal]{IEEEtran}
%
% If IEEEtran.cls has not been installed into the LaTeX system files,
% manually specify the path to it like:
% \documentclass[journal]{../sty/IEEEtran}
\usepackage{graphicx}
\usepackage{float}
\usepackage{subfigure}
\usepackage{booktabs}
\usepackage{CJK}
\usepackage[top=2cm, bottom=2cm, left=2cm, right=2cm]{geometry}
\usepackage{algorithm}
\usepackage{algorithmicx}
\usepackage{algpseudocode}
\usepackage{amsmath}
\usepackage{indentfirst}
\usepackage{amssymb}
\usepackage{multirow}
\usepackage{enumitem}
\newcommand{\CXComment}[1]{\textcolor{red}{#1}}
\usepackage{xcolor}
\usepackage{enumerate}
\usepackage{cite}
\usepackage{ulem}
% Some very useful LaTeX packages include:
% (uncomment the ones you want to load)
% *** MISC UTILITY PACKAGES ***
%
%\usepackage{ifpdf}
% Heiko Oberdiek's ifpdf.sty is very useful if you need conditional
% compilation based on whether the output is pdf or dvi.
% usage:
% \ifpdf
% % pdf code
% \else
% % dvi code
% \fi
% The latest version of ifpdf.sty can be obtained from:
% http://www.ctan.org/pkg/ifpdf
% Also, note that IEEEtran.cls V1.7 and later provides a builtin
% \ifCLASSINFOpdf conditional that works the same way.
% When switching from latex to pdflatex and vice-versa, the compiler may
% have to be run twice to clear warning/error messages.
% *** CITATION PACKAGES ***
%
%\usepackage{cite}
% cite.sty was written by Donald Arseneau
% V1.6 and later of IEEEtran pre-defines the format of the cite.sty package
% \cite{} output to follow that of the IEEE. Loading the cite package will
% result in citation numbers being automatically sorted and properly
% "compressed/ranged". e.g., [1], [9], [2], [7], [5], [6] without using
% cite.sty will become [1], [2], [5]--[7], [9] using cite.sty. cite.sty's
% \cite will automatically add leading space, if needed. Use cite.sty's
% noadjust option (cite.sty V3.8 and later) if you want to turn this off
% such as if a citation ever needs to be enclosed in parenthesis.
% cite.sty is already installed on most LaTeX systems. Be sure and use
% version 5.0 (2009-03-20) and later if using hyperref.sty.
% The latest version can be obtained at:
% http://www.ctan.org/pkg/cite
% The documentation is contained in the cite.sty file itself.
% *** GRAPHICS RELATED PACKAGES ***
%
\ifCLASSINFOpdf
% \usepackage[pdftex]{graphicx}
% declare the path(s) where your graphic files are
% \graphicspath{{../pdf/}{../jpeg/}}
% and their extensions so you won't have to specify these with
% every instance of \includegraphics
% \DeclareGraphicsExtensions{.pdf,.jpeg,.png}
\else
% or other class option (dvipsone, dvipdf, if not using dvips). graphicx
% will default to the driver specified in the system graphics.cfg if no
% driver is specified.
% \usepackage[dvips]{graphicx}
% declare the path(s) where your graphic files are
% \graphicspath{{../eps/}}
% and their extensions so you won't have to specify these with
% every instance of \includegraphics
% \DeclareGraphicsExtensions{.eps}
\fi
% graphicx was written by David Carlisle and Sebastian Rahtz. It is
% required if you want graphics, photos, etc. graphicx.sty is already
% installed on most LaTeX systems. The latest version and documentation
% can be obtained at:
% http://www.ctan.org/pkg/graphicx
% Another good source of documentation is "Using Imported Graphics in
% LaTeX2e" by Keith Reckdahl which can be found at:
% http://www.ctan.org/pkg/epslatex
%
% latex, and pdflatex in dvi mode, support graphics in encapsulated
% postscript (.eps) format. pdflatex in pdf mode supports graphics
% in .pdf, .jpeg, .png and .mps (metapost) formats. Users should ensure
% that all non-photo figures use a vector format (.eps, .pdf, .mps) and
% not a bitmapped formats (.jpeg, .png). The IEEE frowns on bitmapped formats
% which can result in "jaggedy"/blurry rendering of lines and letters as
% well as large increases in file sizes.
%
% You can find documentation about the pdfTeX application at:
% http://www.tug.org/applications/pdftex
% *** MATH PACKAGES ***
%
%\usepackage{amsmath}
% A popular package from the American Mathematical Society that provides
% many useful and powerful commands for dealing with mathematics.
%
% Note that the amsmath package sets \interdisplaylinepenalty to 10000
% thus preventing page breaks from occurring within multiline equations. Use:
%\interdisplaylinepenalty=2500
% after loading amsmath to restore such page breaks as IEEEtran.cls normally
% does. amsmath.sty is already installed on most LaTeX systems. The latest
% version and documentation can be obtained at:
% http://www.ctan.org/pkg/amsmath
% *** SPECIALIZED LIST PACKAGES ***
%
%\usepackage{algorithmic}
% algorithmic.sty was written by Peter Williams and Rogerio Brito.
% This package provides an algorithmic environment fo describing algorithms.
% You can use the algorithmic environment in-text or within a figure
% environment to provide for a floating algorithm. Do NOT use the algorithm
% floating environment provided by algorithm.sty (by the same authors) or
% algorithm2e.sty (by Christophe Fiorio) as the IEEE does not use dedicated
% algorithm float types and packages that provide these will not provide
% correct IEEE style captions. The latest version and documentation of
% algorithmic.sty can be obtained at:
% http://www.ctan.org/pkg/algorithms
% Also of interest may be the (relatively newer and more customizable)
% algorithmicx.sty package by Szasz Janos:
% http://www.ctan.org/pkg/algorithmicx
% *** ALIGNMENT PACKAGES ***
%
%\usepackage{array}
% Frank Mittelbach's and David Carlisle's array.sty patches and improves
% the standard LaTeX2e array and tabular environments to provide better
% appearance and additional user controls. As the default LaTeX2e table
% generation code is lacking to the point of almost being broken with
% respect to the quality of the end results, all users are strongly
% advised to use an enhanced (at the very least that provided by array.sty)
% set of table tools. array.sty is already installed on most systems. The
% latest version and documentation can be obtained at:
% http://www.ctan.org/pkg/array
% IEEEtran contains the IEEEeqnarray family of commands that can be used to
% generate multiline equations as well as matrices, tables, etc., of high
% quality.
% *** SUBFIGURE PACKAGES ***
%\ifCLASSOPTIONcompsoc
% \usepackage[caption=false,font=normalsize,labelfont=sf,textfont=sf]{subfig}
%\else
% \usepackage[caption=false,font=footnotesize]{subfig}
%\fi
% subfig.sty, written by Steven Douglas Cochran, is the modern replacement
% for subfigure.sty, the latter of which is no longer maintained and is
% incompatible with some LaTeX packages including fixltx2e. However,
% subfig.sty requires and automatically loads Axel Sommerfeldt's caption.sty
% which will override IEEEtran.cls' handling of captions and this will result
% in non-IEEE style figure/table captions. To prevent this problem, be sure
% and invoke subfig.sty's "caption=false" package option (available since
% subfig.sty version 1.3, 2005/06/28) as this is will preserve IEEEtran.cls
% handling of captions.
% Note that the Computer Society format requires a larger sans serif font
% than the serif footnote size font used in traditional IEEE formatting
% and thus the need to invoke different subfig.sty package options depending
% on whether compsoc mode has been enabled.
%
% The latest version and documentation of subfig.sty can be obtained at:
% http://www.ctan.org/pkg/subfig
% *** FLOAT PACKAGES ***
%
%\usepackage{fixltx2e}
% fixltx2e, the successor to the earlier fix2col.sty, was written by
% Frank Mittelbach and David Carlisle. This package corrects a few problems
% in the LaTeX2e kernel, the most notable of which is that in current
% LaTeX2e releases, the ordering of single and double column floats is not
% guaranteed to be preserved. Thus, an unpatched LaTeX2e can allow a
% single column figure to be placed prior to an earlier double column
% figure.
% Be aware that LaTeX2e kernels dated 2015 and later have fixltx2e.sty's
% corrections already built into the system in which case a warning will
% be issued if an attempt is made to load fixltx2e.sty as it is no longer
% needed.
% The latest version and documentation can be found at:
% http://www.ctan.org/pkg/fixltx2e
%\usepackage{stfloats}
% stfloats.sty was written by Sigitas Tolusis. This package gives LaTeX2e
% the ability to do double column floats at the bottom of the page as well
% as the top. (e.g., "\begin{figure*}[!b]" is not normally possible in
% LaTeX2e). It also provides a command:
%\fnbelowfloat
% to enable the placement of footnotes below bottom floats (the standard
% LaTeX2e kernel puts them above bottom floats). This is an invasive package
% which rewrites many portions of the LaTeX2e float routines. It may not work
% with other packages that modify the LaTeX2e float routines. The latest
% version and documentation can be obtained at:
% http://www.ctan.org/pkg/stfloats
% Do not use the stfloats baselinefloat ability as the IEEE does not allow
% \baselineskip to stretch. Authors submitting work to the IEEE should note
% that the IEEE rarely uses double column equations and that authors should try
% to avoid such use. Do not be tempted to use the cuted.sty or midfloat.sty
% packages (also by Sigitas Tolusis) as the IEEE does not format its papers in
% such ways.
% Do not attempt to use stfloats with fixltx2e as they are incompatible.
% Instead, use Morten Hogholm'a dblfloatfix which combines the features
% of both fixltx2e and stfloats:
%
% \usepackage{dblfloatfix}
% The latest version can be found at:
% http://www.ctan.org/pkg/dblfloatfix
%\ifCLASSOPTIONcaptionsoff
% \usepackage[nomarkers]{endfloat}
% \let\MYoriglatexcaption\caption
% \renewcommand{\caption}[2][\relax]{\MYoriglatexcaption[#2]{#2}}
%\fi
% endfloat.sty was written by James Darrell McCauley, Jeff Goldberg and
% Axel Sommerfeldt. This package may be useful when used in conjunction with
% IEEEtran.cls' captionsoff option. Some IEEE journals/societies require that
% submissions have lists of figures/tables at the end of the paper and that
% figures/tables without any captions are placed on a page by themselves at
% the end of the document. If needed, the draftcls IEEEtran class option or
% \CLASSINPUTbaselinestretch interface can be used to increase the line
% spacing as well. Be sure and use the nomarkers option of endfloat to
% prevent endfloat from "marking" where the figures would have been placed
% in the text. The two hack lines of code above are a slight modification of
% that suggested by in the endfloat docs (section 8.4.1) to ensure that
% the full captions always appear in the list of figures/tables - even if
% the user used the short optional argument of \caption[]{}.
% IEEE papers do not typically make use of \caption[]'s optional argument,
% so this should not be an issue. A similar trick can be used to disable
% captions of packages such as subfig.sty that lack options to turn off
% the subcaptions:
% For subfig.sty:
% \let\MYorigsubfloat\subfloat
% \renewcommand{\subfloat}[2][\relax]{\MYorigsubfloat[]{#2}}
% However, the above trick will not work if both optional arguments of
% the \subfloat command are used. Furthermore, there needs to be a
% description of each subfigure *somewhere* and endfloat does not add
% subfigure captions to its list of figures. Thus, the best approach is to
% avoid the use of subfigure captions (many IEEE journals avoid them anyway)
% and instead reference/explain all the subfigures within the main caption.
% The latest version of endfloat.sty and its documentation can obtained at:
% http://www.ctan.org/pkg/endfloat
%
% The IEEEtran \ifCLASSOPTIONcaptionsoff conditional can also be used
% later in the document, say, to conditionally put the References on a
% page by themselves.
% *** PDF, URL AND HYPERLINK PACKAGES ***
%
%\usepackage{url}
% url.sty was written by Donald Arseneau. It provides better support for
% handling and breaking URLs. url.sty is already installed on most LaTeX
% systems. The latest version and documentation can be obtained at:
% http://www.ctan.org/pkg/url
% Basically, \url{my_url_here}.
% *** Do not adjust lengths that control margins, column widths, etc. ***
% *** Do not use packages that alter fonts (such as pslatex). ***
% There should be no need to do such things with IEEEtran.cls V1.6 and later.
% (Unless specifically asked to do so by the journal or conference you plan
% to submit to, of course. )
% correct bad hyphenation here
\hyphenation{op-tical net-works semi-conduc-tor}
\begin{document}
%
% paper title
% Titles are generally capitalized except for words such as a, an, and, as,
% at, but, by, for, in, nor, of, on, or, the, to and up, which are usually
% not capitalized unless they are the first or last word of the title.
% Linebreaks \\ can be used within to get better formatting as desired.
% Do not put math or special symbols in the title.
\title{High-order Spatial Interactions Enhanced Lightweight Model for Optical Remote Sensing Image-based Small Ship Detection}
%
%
% author names and IEEE memberships
% note positions of commas and nonbreaking spaces ( ~ ) LaTeX will not break
% a structure at a ~ so this keeps an author's name from being broken across
% two lines.
% use \thanks{} to gain access to the first footnote area
% a separate \thanks must be used for each paragraph as LaTeX2e's \thanks
% was not built to handle multiple paragraphs
%
\author{Yifan Yin,
Xu Cheng,~\IEEEmembership{Member,~IEEE,}
Fan Shi,
Xiufeng Liu,
Huan Huo,
and~Shengyong Chen,~\IEEEmembership{Senior Member,~IEEE}% <-this % stops a space
\thanks{Yifan Yin, Xu Cheng, Fan Shi, and Shengyong Chen are with the School of Computer Science and Engineering, Tianjin University of Technology, Tianjin, China, 300386.}% <-this % stops a space
\thanks{Xiufeng Liu is with the Department of Technology, Management, and Economics, Technical University of Denmark, Produktionstorvet, Denmark, 2800.}% <-this % stops a space
\thanks{Huan Huo is with the School of Computer Science, the University of Technology Sydney, Sydney, Australia, 9220.}}
% note the % following the last \IEEEmembership and also \thanks -
% these prevent an unwanted space from occurring between the last author name
% and the end of the author line. i.e., if you had this:
%
% \author{....lastname \thanks{...} \thanks{...} }
% ^------------^------------^----Do not want these spaces!
%
% a space would be appended to the last name and could cause every name on that
% line to be shifted left slightly. This is one of those "LaTeX things". For
% instance, "\textbf{A} \textbf{B}" will typeset as "A B" not "AB". To get
% "AB" then you have to do: "\textbf{A}\textbf{B}"
% \thanks is no different in this regard, so shield the last } of each \thanks
% that ends a line with a % and do not let a space in before the next \thanks.
% Spaces after \IEEEmembership other than the last one are OK (and needed) as
% you are supposed to have spaces between the names. For what it is worth,
% this is a minor point as most people would not even notice if the said evil
% space somehow managed to creep in.
% The paper headers
\markboth{IEEE Transactions on Geoscience and Remote Sensing}%
{Shell \MakeLowercase{\textit{et al.}}: Bare Demo of IEEEtran.cls for IEEE Journals}
% The only time the second header will appear is for the odd numbered pages
% after the title page when using the twoside option.
%
% *** Note that you probably will NOT want to include the author's ***
% *** name in the headers of peer review papers. ***
% You can use \ifCLASSOPTIONpeerreview for conditional compilation here if
% you desire.
% If you want to put a publisher's ID mark on the page you can do it like
% this:
%\IEEEpubid{0000--0000/00\$00.00~\copyright~2015 IEEE}
% Remember, if you use this you must call \IEEEpubidadjcol in the second
% column for its text to clear the IEEEpubid mark.
% use for special paper notices
%\IEEEspecialpapernotice{(Invited Paper)}
% make the title area
\maketitle
% As a general rule, do not put math, special symbols or citations
% in the abstract or keywords.
\begin{abstract}
Accurate and reliable optical remote sensing image-based small-ship detection is crucial for maritime surveillance systems, but existing methods often struggle with balancing detection performance and computational complexity. In this paper, we propose a novel lightweight framework called \textit{HSI-ShipDetectionNet} that is based on high-order spatial interactions and is suitable for deployment on resource-limited platforms, such as satellites and unmanned aerial vehicles. HSI-ShipDetectionNet includes a prediction branch specifically for tiny ships and a lightweight hybrid attention block for reduced complexity. Additionally, the use of a high-order spatial interactions module improves advanced feature understanding and modeling ability. Our model is evaluated using the public Kaggle marine ship detection dataset and compared with multiple state-of-the-art models including small object detection models, lightweight detection models, and ship detection models. The results show that HSI-ShipDetectionNet outperforms the other models in terms of recall, and mean average precision (mAP) while being lightweight and suitable for deployment on resource-limited platforms.
\end{abstract}
% Note that keywords are not normally used for peerreview papers.
\begin{IEEEkeywords}
Small ship detection, Optical remote sensing images, Convolutional neural networks, Spatial interaction, Lightweight model.
\end{IEEEkeywords}
% For peer review papers, you can put extra information on the cover
% page as needed:
% \ifCLASSOPTIONpeerreview
% \begin{center} \bfseries EDICS Category: 3-BBND \end{center}
% \fi
%
% For peerreview papers, this IEEEtran command inserts a page break and
% creates the second title. It will be ignored for other modes.
\IEEEpeerreviewmaketitle
\section{Introduction}
\label{Introduction}
\IEEEPARstart{M}{onitoring} the position and behavior of ships plays a critical role in maintaining marine traffic safety and supporting social and economic development. The use of optical remote sensing images provides valuable information for various applications such as fishery management, marine spatial planning, marine casualty investigation, and pollution treatment \cite{eldhuset1996automatic,yin2022enhanced}. However, when the altitude and angle of satellite photography vary, ship targets can have a large scale of variation, so there are a large number of small target ships in the images. The complex sea state can significantly impact the detection performance of small ships. Waves can cause variations in pixel values in the optical image due to the reflection of the sun and skylight off their slopes \cite{kanjir2018vessel}. Additionally, satellites may encounter clouds or sunglint when observing the Earth, which can make it difficult to distinguish ships from the background, even for the naked eye \cite{tian2021image}. Therefore, it is still difficult to accurately locate and recognize small ships from optical remote sensing images.
%In the past decades, small ship detection in optical remote sensing images has been studied extensively. Many traditional methods mainly focus on feature design, which generally includes ship candidate extraction and ship identification \cite{bo2021ship}. There are many basic techniques in the ship candidate extraction stage, such as the statistical threshold segmentation \cite{pegler2003potential,kanjir2014automatic}, visual saliency \cite{xu2016ship}, and local feature descriptor \cite{selvi2011novel}, etc. In the next step of identification operations, the support vector machine (SVM) \cite{xia2011novel} is a frequently adopted method for ship classification. However, the impact of variable weather factors on optical image imaging is uncontrollable, these traditional algorithms may not adapt to complex conditions. In addition, the production and generation of features strongly rely on manual and expert experience, so their generalization ability is poor.
Over the past few decades, there has been a significant amount of research on small ship detection in optical remote sensing images. Traditional methods have mainly focused on feature design, including ship candidate extraction and ship identification \cite{bo2021ship}. Ship candidate extraction techniques such as statistical threshold segmentation \cite{pegler2003potential,kanjir2014automatic}, visual saliency \cite{xu2016ship}, and local feature descriptor \cite{selvi2011novel} have been commonly used. In the identification stage, the support vector machine (SVM) \cite{xia2011novel} has been a frequently adopted method for ship classification. However, traditional methods may not be effective in complex conditions as the impact of variable weather factors on optical image imaging is uncontrollable. Additionally, these algorithms rely heavily on manual and expert experience for feature production and generation, resulting in poor generalization ability.
Recently, the use of convolutional neural networks (CNNs) has greatly improved the accuracy and efficiency of ship detection. However, the continuous downsampling characteristic of CNNs can still present challenges for detecting small ships in optical remote sensing images. One important way to improve the detection accuracy of small objects is to address the issue of multi-scale feature learning. Shallow layers of convolutional neural networks (CNNs) typically have higher resolutions and smaller receptive fields, which are more suitable for detecting small objects \cite{tong2020recent}. Several methods have been developed to make use of these shallow layers for small object detection, including the Single Shot MultiBox Detector (SSD) \cite{liu2016ssd} and the top-down feature pyramid network (FPN) with lateral connections \cite{lin2017feature}. In addition to multi-scale feature learning, the use of contextual information can also be beneficial for improving object detection performance, particularly for small objects with insufficient pixels \cite{tong2020recent}. This is because specific objects often appear in specific environments, such as ships sailing in the sea. Context-based small object detection methods can be divided into two categories: local context modeling \cite{ zagoruyko2016multipath,guan2018scan} and global context modeling \cite{ouyang2015deepid,zhu2015segdeepm,zhu2021tph}.
%In recent years, the success of deep learning methods has promoted the exploration of small object detection. Firstly, an important means to promote the detection accuracy of small objects is to solve the problem of multi-scale feature learning. To be specific, There are higher resolutions and smaller receptive fields in the features in shallow layers of CNN, which are more suitable for detecting small objects\cite{tong2020recent}. Liu et al. 's SSD\cite{liu2016ssd} constructed a pyramidal feature hierarchy that allowed predictions from multiple scales, where the shallow layer was in charge of small objects. Later, a top-down feature pyramid network (FPN) with lateral connections was developed by Lin et al\cite{lin2017feature}. This structure made the shallow spatially-rich features be reinforced by deep semantic-rich features. Subsequently, several FPN-based variants have been designed specifically for the detection of small objects\cite{liang2018small, cao2018feature, min2022attentional}. Chen et al.\cite{chen2020detection} presented an optimized FPN model for small ship detection under complex background. In addition, reasonable use of contextual information is also beneficial to improve object detection performance, particularly for small objects with insufficient pixels\cite{tong2020recent}. This is because specific objects usually appear in a specific environment, such as this study scenario in this paper: ships are sailing in the sea. Context-based small object detection methods can be divided into two categories: local context modeling and global context modeling. Among them, a novel semantic context aware network (SCAN) proposed by Guan et al.\cite{guan2018scan} combined local context information and location information to enhance the detection performance of small objects. In terms of the global context, the Transformer mechanism\cite{vaswani2017attention, dosovitskiy2020image} has gained a lot of attention recently due to its computational approach. Zhu et al.\cite{zhu2021tph} embedded Transformers into the YOLOv5 framework and built a detector TPH-YOLOv5 for small objects, which has achieved excellent results.
Despite the advancements made by CNN-based detection networks in improving the detection performance of small objects, several limitations persist. These limitations include:
\begin{itemize}[leftmargin=*,noitemsep]
\item The utilization of multi-scale feature learning has had a positive impact on the detection accuracy of small ships. However, it has been observed that most existing networks are limited to three scales \cite{chen2021improved, xie2022small}. This is deemed to be insufficient as the shallow features, which are crucial in detecting tiny objects, are not fully utilized. How to take the full utilization of shallow features is challenging.
\item The use of CNN-based models for small object detection has been shown to be effective, but these models often have a high number of parameters and are complex. For example, the TPH-YOLOv5 detector \cite{zhu2021tph} is well-known for its proficiency in small object detection, but it requires 60 million parameters. This complexity can lead to time delays when transmitting data from the platform to ground stations for processing \cite{xu2022lite}. To address this, it is necessary to migrate ship detection models from ground to space-borne platforms. However, hardware resources on such platforms are often limited, such as the NVIDIA Jetson TX2 which only has 8 GB of memory \cite{xu2021board}. This makes it difficult to reduce model complexity while still maintaining accuracy in ship detection. Therefore, finding an optimal balance between model accuracy and complexity is an ongoing research challenge.
\item As the depth of the network layers increases, the high-level features at the end of the backbone exhibit an abundance of combinatorial information. While these higher-level features carry richer semantic information, the location information they convey is ambiguous. This ambiguity can negatively impact the accuracy of small object detection, particularly for objects with insufficient pixels \cite{hu2019sar}, making it challenging to accurately localize and regress small target ships. Additionally, the complexity of the background texture and harsh environmental conditions can weaken the ability of CNNs to extract features of ships, making it difficult to distinguish small ships from their background.
\end{itemize}
Given the limitations of existing methods in small ship detection and the need to balance detection performance with the limited storage space available on satellites, this paper proposes a novel lightweight ship detection framework based on high-order spatial interactions (HSI). The contributions of this study can be summarized as follows:
\begin{itemize}[leftmargin=*,noitemsep]
\item This study proposes an enhanced ship detection network, HSI-ShipDetectionNet, which is designed to be more lightweight and effective for ship detection in optical remote sensing images. Furthermore, the proposed network demonstrates improved accuracy in the localization and identification of small ships.
\item To make the detection model more accurate in detecting tiny ships, we add a predictive branch of tiny ships ($P_{tiny}$). To support this branch, we increase the number of layers in the neck of the detection frame, making the model more sensitive to tiny ships. Then, we design a lightweight hybrid attention block (LHAB) to replace the SE block in GhostNet, which is the backbone of the HSI-ShipDetectionNet, reducing the number of parameters, computations, and storage space required by the model. Finally, A high-order spatial interactions (HSI-Former) module is introduced at the tail of the backbone, extending the interaction between spatial elements to any order and strengthening the model’s ability to understand and process advanced features in deep layers. And in it, we use large convolutional kernels for context modeling to improve the accuracy of ship position regression.
\item We comprehensively evaluate the proposed ship detection framework using optical satellite remote sensing images. The performance of the proposed model is compared with that of state-of-the-art small object detection models, lightweight detection models, and ship detection models. The experimental results indicate that the proposed HSI-ShipDetectionNet demonstrates remarkable performance in detecting small ships under diverse sea conditions, as well as under a wide range of altitude and angle variations. Furthermore, the lightweight nature of the proposed model makes it highly suitable for deployment on resource-constrained satellite platforms.
\end{itemize}
The remainder of this article is organized as follows: Section \ref{relatedwork} reviews related work on this topic. Section \ref{method} outlines the framework of our discussed methodology. Section \ref{experiments} describes the experimental results and analysis, and Section \ref{conclusion} concludes the whole study.
\section{Related Work}
\label{relatedwork}
\subsection{Methods for Small Ship Detection}
%In the Introduction, we have already mentioned two main methods for small object detection: multi-scale feature fusion and context-based learning. In this subsection, we will briefly review the literature on enhancing small ship detection performance in the field of ship detection.
%Accurate and dependable detection for small ships is crucial for maritime surveillance systems. In recent years, extensive efforts have been undertaken to detect small ships. For instance, Wu et al.\cite{wu2020coarse} presented a multi-scale detection strategy to achieve varying degrees of refinement. After that, to prevent the pixel loss of small ships, they embed FPN into the coarse-to-fine ship detection network (CF-SDN), which can not only obtain higher resolution in shallow feature maps but also learn more semantic information in deep feature maps. In addition, Xie et al.\cite{xie2022small} argued that traditional FPN cannot adaptively adjust weights according to the importance of features. So they introduced the adaptive feature enhancement (AFE) module into FPN, which uses shallow features with rich spatial information to adaptively reinforce the locations of the deep features. Wang et al.\cite{wang2022yolo} Developed a ship detection model based on YOLOX, in which multi-scale convolution (MSC) was designed for feature fusion and a feature transformer module (FTM) was proposed for context modeling. Jin et al.\cite{jin2020patch} exploited another approach for context modeling, which is to input patches containing targets and surroundings into CNN to improve the detection results of small ships. However, the above networks are complex, so it is limited in some resource-scarce application scenarios.
Accurate and dependable detection of small ships is crucial for maritime surveillance systems. In recent years, there have been numerous efforts to improve the performance of small ship detection.
With the development of deep learning, the use of convolutional neural networks (CNNs) for ship detection has become mainstream. For example, Wu et al. \cite{wu2020coarse} proposed a multi-scale detection strategy that uses a coarse-to-fine ship detection network (CF-SDN) with a feature pyramid network (FPN) to improve the resolution and semantic information of shallow and deep feature maps, respectively. Xie et al. \cite{xie2022small} introduced an adaptive feature enhancement (AFE) module into FPN to adaptively reinforce the locations of deep ship features based on shallow features with rich spatial information. Wang et al. \cite{wang2022yolo} developed a ship detection model based on YOLOX that incorporates a multi-scale convolution (MSC) for feature fusion and a feature transformer module (FTM) for context modeling. Jin et al. \cite{jin2020patch} input patches containing targets and surroundings into a CNN to improve small ship detection results. Tian et al. \cite{tian2021image} proposed an image enhancement module base on generative adversarial network (GAN), and introduced the receptive field expansion module to improve the capability to extract features from target ships of different sizes.
Despite the remarkable detection performance demonstrated by existing ship detection models, these models are often characterized by large and complex network architectures, as evidenced by the substantial number of parameters and computational demands. This presents a significant challenge for resource-constrained applications, where the available hardware resources are limited. To overcome this limitation, we design a lightweight attention block and construct a lightweight ship detection framework, which reduces the number of parameters, computations, and storage space required by the model.
%Recently, the use of high-order spatial interactions (HSI) has been proposed as a promising approach for object detection in images \cite{liu2019high}. HSI models capture the dependencies between the features of different pixels in the image, allowing for the modeling of complex relationships and the incorporation of contextual information \cite{liu2018high}. In this work, we propose the use of HSI for small ship detection in optical remote sensing images as a lightweight alternative to traditional approaches.
\subsection{Methods for Lightweight CNNs}
%To deploy the network model to some resource-limited devices, it is necessary to conduct lightweight design of the networks, mainly from the aspects of parameter quantity and calculated amount\cite{zhou2020review}. SqueezeNet\cite{iandola2016squeezenet} mainly emphasized reducing the feature map with a $1\times1$ convolution kernel to reduce the network parameters. The main core of the MobileNet series\cite{howard2017mobilenets,sandler2018mobilenetv2,howard2019searching} is the stacking of depthwise separable convolution. Depthwise separable convolution is a form of factorized convolution that factorize a standard convolution into a depthwise convolution and a pointwise convolution. This operation can reduce the number of parameters and computations of the network to $\frac{1}{N_{output}}{+} \frac{1}{k^2}$. Where $N_{output}$ represents the number of output channels and $k$ is the size of the convolution kernel. ShuffleNet's\cite{zhang2018shufflenet,ma2018shufflenet} main contribution is to replace the pointwise convolution with pointwise group convolution and then perform channel shuffle. This can not only further lessen the parameters but also make up for the disadvantages of group convolution, so that the information could flow between each group. GhostNet\cite{han2020ghostnet} proposed the concept of cheap operation, which is a cost-efficient way to embrace abundant and even redundant information.
Lightweight design of CNNs is crucial for deploying models to resource-limited devices such as satellites, as it helps to reduce the number of parameters and computational requirements. A number of approaches have been proposed in the literature to achieve this goal, including SqueezeNet \cite{iandola2016squeezenet}, which reduces the number of parameters by using $1\times1$ convolution kernels to decrease the size of the feature maps; the MobileNet series \cite{howard2017mobilenets, sandler2018mobilenetv2, howard2019searching}, which uses depthwise separable convolution to factorize standard convolution into a depthwise convolution and a pointwise convolution, reducing the number of parameters and computational requirements; ShuffleNet \cite{zhang2018shufflenet, ma2018shufflenet}, which replaces pointwise convolution with pointwise group convolution and performs channel shuffle to further reduce the number of parameters and address the disadvantages of group convolution; and GhostNet \cite{han2020ghostnet}, which embraces abundant and redundant information through cheap operations as a cost-efficient way to improve network performance.
%In the practical application of the ship detection field, it is difficult to balance the detection performance and computational complexity of the model. To solve this challenge, Li et al.\cite{li2020lightweight} optimized the backbone of YOLOv3 using the dense connection, and introduced spatial separation convolution to replace standard convolution in FPN, thus greatly reducing parameters. Jiang et al.\cite{jiang2021high} developed YOLO-V4-light by lowering the number of convolutional layers in CSPDarkNet53. Liu et al.\cite{liu2022multi} also improved the network based on YOLOv4, substituting the original backbone with MobileNetv2, which widely cut down the complexity of the ship detection model. Zheng et al.\cite{zheng2022fast} used BN scaling factor $\gamma$ to compress the YOLOv5 network. Experimental results show that YOLOv5 has higher detection accuracy and shorter computational time compared to other object detection models.
In the field of ship detection, it can be challenging to balance the performance and computational complexity of the model. To address this issue, Li et al. \cite{li2020lightweight} optimized the backbone of YOLOv3 using dense connections and introduced spatial separation convolution to replace standard convolution in FPN, resulting in a significant reduction in parameters. Jiang et al. \cite{jiang2021high} developed YOLO-V4-light by reducing the number of convolutional layers in CSPDarkNet53. Liu et al. \cite{liu2022multi} also improved upon YOLOv4 by substituting the original backbone with MobileNetv2, significantly reducing the complexity of the ship detection model. Zheng et al. \cite{zheng2022fast} used BN scaling factor $\gamma$ to compress the YOLOv5 network, achieving higher detection accuracy and shorter computational time compared to other object detection models.
\subsection{Methods for Attention Mechanism}
%Attention has become one of the most crucial concepts in the computer vision field, which can dramatically improve the performance of the network\cite{niu2021review}. Channel attention allows networks to model the dependencies between the channels of their convolutional features. Squeeze-and-excitation (SE) network\cite{hu2018squeeze} proposed by Hu et al. could adaptively recalibrate channel-wise features, using global information to selectively highlight vital features. Based on SE, Wang et al.\cite{2020ECA} designed an efficient channel attention (ECA) that can be implemented by 1D convolution, which proved to be more efficient and effective. In addition, there is spatial attention, which is designed to find the position in the image that should be focused on. For example, CCNet\cite{huang2019ccnet} was presented to capture full-image contextual information employing criss-cross attention. Furthermore, Convolutional Block Attention Module (CBAM)\cite{woo2018cbam} blended channel attention and spatial attention, emphasizing important features along those two dimensions.
Attention mechanisms have become a key concept in the field of computer vision, with the ability to significantly improve the performance of networks \cite{niu2021review}. Channel attention allows networks to model dependencies between the channels of their convolutional features, such as in the Squeeze-and-excitation (SE) network \cite{hu2018squeeze}, which adaptively recalibrates channel-wise features using global information to selectively highlight important features. Wang et al. \cite{2020ECA} further developed this concept with the efficient channel attention (ECA) module, which can be implemented using 1D convolution and has been shown to be more efficient and effective. Spatial attention, on the other hand, focuses on identifying specific positions in the image that should be emphasized, such as in CCNet \cite{huang2019ccnet}, which captures full-image contextual information using criss-cross attention. The Convolutional Block Attention Module (CBAM) \cite{woo2018cbam} combines channel and spatial attention, emphasizing important features in both dimensions.
%The Transformer proposed by Vaswani et al.\cite{vaswani2017attention} is an important milestone in the attention mechanism, and the Transformer model applied in the field of computer vision is called Vision Transformer (ViT)\cite{dosovitskiy2020image}. The core idea is to use the Self-Attention mechanism to dynamically generate weights to establish long-range dependencies. By matrix multiplication between queries, keys, and values in self-attention, the interaction of two spatial is performed. However, Rao et al.\cite{rao2022hornet} proved that higher order spatial interactions could improve the modeling ability of the visual model, so they proposed a gated convolution that could expand the spatial interactions to any order. In this work, we will develop a YOLOv5-based lightweight ship detection framework for small ships, according to the following aspects, i.e., extending FPN by adding a predictive branch of tiny ships, exploiting the lightweight hybrid attention block (LHAB), and introducing the high-order spatial interactions module (HSI-Former), leading to more accurate and reliable ship detection in the surveillance system. The ablation experiments and comparison experiments will be implemented to demonstrate the superior performance of our model.
The Transformer model, proposed by Vaswani et al. \cite{vaswani2017attention}, has been a major milestone in the development of attention mechanisms, and its application to the field of computer vision is known as the Vision Transformer (ViT) \cite{dosovitskiy2020image}. The core idea of the Transformer is to use self-attention to dynamically generate weights that establish long-range dependencies. Self-attention achieves this through matrix multiplication between queries, keys, and values, allowing for the interaction of two spatial elements. However, it has been noted that the Transformer architecture is limited in its capability to model higher-order spatial interactions, which can potentially enhance the overall visual modeling performance\cite{rao2022hornet}. In this work, we propose a novel lightweight ship detection framework for small ships that includes the following elements: an extension of FPN through the addition of a predictive branch for tiny ships, the use of the lightweight hybrid attention block (LHAB), and the introduction of the high-order spatial interactions (HSI-Former) module, resulting in more accurate and reliable ship detection in surveillance systems. Ablation and comparison experiments will be conducted to demonstrate the superior performance of our model.
\section{Methodology}
\label{method}
\subsection{Overview}
%As shown in Fig.~\ref{network}, the overall process of the proposed lightweight HSI-ShipDetectionNet for small ship detection contains three major parts: the Backbone, the HSI-Former model, and the Neck. For the input optical remote sensing images, the backbone is first responsible for extracting the detailed features of the ship. To address the challenge of the small ship detection mentioned in Section \ref{Introduction}, a predictive branch for tiny ships is added to the shallow layer of the backbone, as explained in detail in Section \ref{The Predictive Branch of Tiny Ships}. In addition, To further reduce the complexity of the model, we improve the Ghost bottleneck in GhostNet. A new Lightweight Hybrid Attention Block (LHAB) is designed to replace the SE block\cite{hu2018squeeze} to obtain a new LHAB-Gbneck, resulting in a significant reduction in the number of parameters, the computational effort, and the occupied storage space by the network. The details are illustrated in Section \ref{LHAB-GhostCNN}. Furthermore, we introduce a high-order spatial interactions (HSI-Former) module at the tail of the backbone, which strengthens the contextual learning and modeling ability of the advanced features in the deep layers. We give details of this in Section \ref{High-Order Spatial Interaction Mechanism}. The data then moves on to the next stage - the neck layer - where the fusion of features takes place. Eventually, four different output heads are set up to predict tiny, small, medium, and large ship targets respectively.
The proposed lightweight HSI-ShipDetectionNet for small ship detection, as depicted in Fig.~\ref{network}, consists of three key components: the Backbone, the HSI-Former module, and the Neck. The input optical remote sensing images undergo processing in the backbone, which extracts the detailed features of the ship. To address the challenge of small ship detection, a predictive branch specifically designed for tiny ships is added to the shallow layer of the backbone, as discussed in detail in Section \ref{The Predictive Branch of Tiny Ships}. To further reduce the complexity of the model, the Ghost bottleneck in GhostNet has been improved with the implementation of a new Lightweight Hybrid Attention Block (LHAB), which replaces the SE block\cite{hu2018squeeze}. This results in a LHAB-Gbneck with a reduced number of parameters, computational effort, and occupied storage space, as explained in Section \ref{LHAB-GhostCNN}. In addition, the HSI-Former module, which is designed to reinforce contextual learning and modeling capability of advanced features in deep layers, is introduced at the tail of the backbone. The function and implementation of the HSI-Former module are detailed in Section \ref{High-Order Spatial Interaction Mechanism}. Finally, the neck layer fuses the features, and four separate output heads are employed to predict tiny, small, medium, and large ship targets, respectively.
%The overall network structure of the proposed HSI-ShipDetectionNet is illustrated in Fig.~\ref{network}. The model contains three main parts: the Backbone, the HSI-Former module, and the Neck. The Backbone is responsible for extracting detailed features of ships from the input remote sensing images. To improve the detection performance of small ships, a predictive branch for tiny ships is added to the shallow layer of the Backbone. To reduce the complexity of the model, the Ghost bottleneck in GhostNet is improved by introducing a Lightweight Hybrid Attention Block (LHAB) to replace the SE block \cite{hu2018squeeze}, resulting in a significant reduction in the number of parameters, computational effort, and storage space required by the network. The HSI-Former module is added to the tail of the Backbone to enhance the contextual learning and modeling of advanced features in the deep layers. The Neck layer then performs feature fusion, and four output heads are used to predict tiny, small, medium, and large ships respectively. In the following subsections, we will describe each module in detail.
\begin{figure*}[ht]
\centering
\includegraphics[width=1\textwidth]{Images/Network.pdf}
\caption{Overview of the proposed HSI-ShipDetectionNet for small ship detection in optical remote sensing images. In the \textbf{Backbone}, a predictive branch is added to the shallow layer specifically for detecting tiny ships. The \textbf{Lightweight Hybrid Attention Block (LHAB)} in \textbf{LHAB-Gbneck} is designed, resulting in a significant reduction in the number of parameters, computational effort, and storage space required by the network. The \textbf{HSI-Former module} is added to the end of the Backbone to enhance the contextual learning and modeling of advanced features in the deep layers. The \textbf{Neck} layer then performs feature fusion, and four output heads are used to predict tiny, small, medium, and large ships respectively.}
\label{network}
\end{figure*}
\subsection{The Predictive Branch of Tiny Ships}
\label{The Predictive Branch of Tiny Ships}
%With the deepening of the backbone, the features of small ships are continuously down-sampled by the convolutional layers, and the resolution of the output small targets gets smaller and even lost \cite{wu2020coarse}. This results in low detection accuracy for small ships. And unfortunately, satellite images contain a large number of small ships. Therefore, if the features are extracted before continuous downsampling, that is, the information on small ships in shallow feature maps is captured, it will be easier to obtain the feature and location information. According to the above analysis, as shown in Figure x, we add a branch predicting tiny ships in stage 1 of the backbone and name it $P_{tiny}$, which is more sensitive to the tiny ships. Meanwhile, the number of layers of the PANet in the neck of the detection frame is also increased, enhancing the feature fusion effect of tiny ships. This structure gradually fuses shallow features to deep layers, so that the feature maps of different sizes contain both the semantic information and feature information of ships, which ensures the detection accuracy of the ships with different scales, especially for tiny ships.
The problem of low detection accuracy for small ships in satellite imagery is a well-known issue. This is due to the continuous down-sampling of features by the convolutional layers in the backbone network, which results in the loss of resolution and information for small ships. Small ships are often present in satellite images, making it crucial to address this problem to improve overall detection accuracy. To address this issue, we propose adding a branch that predicts tiny ships in stage 1 of the backbone, as shown in Fig.~\ref{network}. This branch, named $P_{tiny}$, is specifically designed to be more sensitive to tiny ships. Additionally, the number of layers in the PANet in the neck of the detection frame is increased to enhance the feature fusion effect for tiny ships. This structure gradually fuses shallow features with deep layers, ensuring that the feature maps of different sizes contain both semantic information and feature information of ships. This ultimately ensures the detection accuracy of ships with different scales, particularly for tiny ships. By extracting features before the continuous downsampling process, the detection accuracy of small ships is expected to be improved.
%Since we add another branch to the original three branches of YOLOv5, we also add another set of anchors for tiny ships accordingly, that is, four groups in total. Instead of using the COCO dataset in the original YOLOv5, we use clustering to generate anchors specifically for ship sizes in our dataset. This makes the regression of the anchors more accurate. As mentioned in \cite{redmon2017yolo9000}, the effect of using Euclidean distance is not as good as that of 1-IOU, we choose 1-IOU as the distance for the clustering. The sizes of the four groups of anchors are as follows: (7,16, 10,9, 18,7), (16,15, 20,27, 34,16), (37,30, 60,21, 26,58) and (63,34, 45,54, 66,57), each of which has three different sizes of anchors, twelve in total.
Along with this new branch, we also add an additional set of anchors specifically tailored for tiny ships based on the original three groups of anchors of YOLOv5, resulting in a total of four groups of anchors. Instead of using the anchors generated by COCO dataset as in the original YOLOv5, we employ clustering to generate new anchors specifically for ship sizes in our dataset. This makes the regression of the anchors more accurate. As per the research in \cite{redmon2017yolo9000}, we have chosen 1-IOU as the distance for the clustering instead of Euclidean distance for better results. The sizes of the four groups of anchors are as follows: (7,16, 10,9, 18,7), (16,15, 20,27, 34,16), (37,30, 60,21, 26,58) and (63,34, 45,54, 66,57), each of which has three different sizes of anchors, resulting in a total of twelve anchors.
\subsection{LHAB-GhostCNN}
\label{LHAB-GhostCNN}
%We believe that GhostNet is well-suited as the backbone of the detection network for small ships. This is due to the core idea of GhostNet - "cheap operation". The authors of GhostNet visualized the feature maps generated by the first residual group in ResNet-50 and found that some of these feature maps are very similar, which indicates that there is abundant and redundant information in the feature maps. They considered that it is this abundant and redundant information that guarantees a comprehensive understanding of the input data. So rather than discarding the redundant feature maps, they were inclined to accept them in a cost-efficient way known as a cheap operation.
%Small ships occupy fewer pixel units, so the information on these is extremely valuable. Therefore, removing some redundant information to reduce the complexity of the network is not a good way for small ship detection. However, GhostNet chooses to embrace them in a cost-effective way, which is conducive to small targets. Thus, We select GhostNet as the backbone of the lightweight ship detector and further simplify it.
We propose using GhostNet as the backbone of the detection network for small ships. The core idea behind GhostNet is "cheap operation" which is well-suited for small ship detection. The authors of GhostNet found that some of the feature maps generated by the first residual group in ResNet-50 were very similar, indicating that there was abundant and redundant information in the feature maps. Rather than discarding these redundant feature maps, they chose to accept them in a cost-efficient way.
Small ships occupy fewer pixel units, making the information about them extremely valuable. Removing redundant information to reduce the complexity of the network is not a good approach for small ship detection. However, GhostNet's approach of embracing redundant information in a cost-effective way is beneficial for small target detection. Therefore, we have selected GhostNet as the backbone of our lightweight ship detector and further simplified it. We name this architecture as \textit{LHAB-GhostCNN}.
\subsubsection{Ghost Module}
%The Ghost module is a key component of our proposed LHAB-GhostCNN architecture for small ship detection. The Ghost module is designed to maintain the same number of feature maps as a standard convolution while reducing the number of parameters and computational effort. Assume that the input feature maps are $C$, and after standard convolution, it outputs $D$. The Ghost module can also obtain $D$ feature maps while decreasing the number of parameters and computations without avoiding redundant information. The process is defined as follows.
The Ghost module is a crucial element of the proposed LHAB-GhostCNN architecture for small ship detection. Its purpose is to maintain the same number of feature maps as a standard convolution while reducing the number of parameters and computational effort. Specifically, when the input feature maps are $C$ and the output feature maps after standard convolution are $D$, the Ghost module can also produce $D$ feature maps while minimizing the number of parameters and computations, without compromising redundant information. The process can be defined as follows.
For the input feature $X\in\mathbb{R}^{H\times W\times C}$, the $m$ intrinsic feature maps are first generated by a standard convolution, represented by the set $Y_1$:
\begin{equation}
Y_1=Conv\left(X\right), \quad Y_1\in\mathbb{R}^{H'\times W'\times m}
\end{equation}
%where $m\leq D$. To obtain the desired $D$ feature maps, each intrinsic feature map of $Y_1$ is then performed with $s$ cheap operations (implemented by depthwise separable convolution, that is DW-Conv) to gain $m×s$ ghost feature maps $Y_2$:
where $m\leq D$. To obtain the desired $D$ feature maps, each of the $m$ intrinsic feature maps in $Y_1$ undergoes $s$ cheap operations, implemented through depthwise separable convolution (DW-Conv), resulting in $m \times s$ ghost feature maps $Y_2$:
\begin{equation}
\begin{split}
&Y_2=\Phi\left(Y_1\right): y_{ij}=DW\_Conv_{ij}\left(y_i\right),\\
&\forall i=1,\dotsi,m,\quad j=1,\dotsi,s
\end{split}
\end{equation}
where $y_i$ represents the $i$-th intrinsic feature map in $Y_1$, and the $j$-th feature map $y_{ij}$ is generated by the $j$-th linear operation $DW-Conv_{ij}$. As a result, these $m$ intrinsic feature maps can eventually generate $m×s$ feature maps, that is, $Y_2\in\mathbb{R}^{H'\times W'\times ms}$. The final output of the Ghost module is the concatenation of $Y_1$ and $Y_2$:
\begin{equation}
Y_{out}=Y_1\oplus Y_2
\end{equation}
%Through this operation, the same number of $D$ feature maps as the standard convolution is obtained, so the dimension of $Y_{out}$ is $m+ms=D$.
By employing the Ghost module, $D$ feature maps can be obtained while maintaining the same number of feature maps as a standard convolution. Consequently, the output feature maps $Y_{out}$ have a dimension of $m+ms=D$.
\textbf{Analysis of complexities.} We define $r_F$ as the speed-up ratio of FLOPs of the Ghost module to FLOPs of the standard convolution:
\begin{equation}
\begin{split}
r_F&= \frac{k\cdot k \cdot C\cdot m\cdot H' \cdot W'+ d\cdot d \cdot m\cdot s\cdot H' \cdot W'}{k\cdot k \cdot C\cdot D\cdot H' \cdot W'}\\
&= \frac{C\cdot m+ m\cdot s\cdot9}{C\cdot m \cdot\left(1+s \right) }= \frac{C+ s\cdot9}{C\cdot \left(1+s \right) } \approx \frac{1}{1+s}
\end{split}
\end{equation}
where $k=1$ is the standard convolution kernel size, while $d=3$ is the kernel size of each linear operation, and $C\gg s$. Similarly, the compression ratio $r_P$ of the parameters of the Ghost module to the parameters of the standard convolution is:
\begin{equation}
\begin{split}
r_P&= \frac{k\cdot k \cdot C\cdot m+ d\cdot d \cdot m\cdot s}{k\cdot k \cdot C\cdot D}\\
&= \frac{C\cdot m+ m\cdot s\cdot9}{C\cdot m \cdot\left(1+s \right) }= \frac{C+ s\cdot9}{C\cdot \left(1+s \right) } \approx \frac{1}{1+s}
\end{split}
\end{equation}
%In this paper, we set $s$ to 1. Thus, the Ghost Module can reduce the size of parameters and the computational effort of the network to half.
In this paper, we set the value of $s$ to 1. As a result, the Ghost module can effectively reduce the number of parameters and the computational effort of the network by half.
\subsubsection{LHAB-Gbneck}
\label{LHAB-Gbneck}
Similar to the basic residual block in ResNet \cite{he2016deep}, the Ghost bottleneck with LHAB (LHAB-Gbneck) integrates two Ghost modules and a shortcut, as shown in Fig.~\ref{bottleneck}. The first Ghost module serves as an expansion layer to increase the number of channels, while the second Ghost module reduces the number of channels to match the shortcut connection. The shortcut is connected between the inputs and outputs of these two Ghost modules. When Stride=2, a depthwise separable convolution (DW-Conv) is added after the first Ghost module to reduce the size of the feature maps by half, at this time the shortcut path goes through a downsampling layer to match the size of the feature maps. If LHAB=1, the Lightweight Hybrid Attention Block (LHAB) is selected. Compared with the SE attention \cite{hu2018squeeze} used in the original Ghost bottleneck, LHAB can further reduce the complexity of the network while enhancing the response of key features.
\begin{figure}[ht]
\centering
\includegraphics[width=0.48\textwidth]{Images/bottleneck.pdf}
\caption{LHAB-Gbneck. Stride=1 and Stride=2 go through different branches.}
\label{bottleneck}
\end{figure}
%The SE block, as a channel attention mechanism, ignores the importance of spatial attention and introduces much additional complexity into the model. To balance the performance of the model with its complexity, we propose the Lightweight Hybrid Attention Block (LHAB), a lightweight yet efficient attention block. LHAB is divided into two parts: a channel attention block and a spatial attention block, so it can highlight meaningful information along those two dimensions simultaneously.
The SE block, a widely used channel attention mechanism, has limitations in ignoring spatial attention and adding complexity to the model. To balance the trade-off between model performance and complexity, we propose the Lightweight Hybrid Attention Block (LHAB), which is a lightweight and efficient attention block. LHAB consists of a channel attention block and a spatial attention block, enabling it to highlight significant information in both dimensions simultaneously.
\textbf{Channel attention block.}
The channel attention block is a key component of the LHAB, which aims to capture interdependencies between channels. SENet \cite{hu2018squeeze} employed global average pooling to aggregate channel-wise statistics, but it overlooks the potential of max-pooling in inferring fine channel attention, as pointed out by Woo et al. \cite{woo2018cbam}. Therefore, they proposed to use both average-pooling and max-pooling operations in tandem and generated the channel attention map using a shared network. In contrast, we believe that max-pooled features and average-pooled features each play distinct roles and therefore require dedicated parameters to store unique feature information. Therefore, we do not use shared parameters and instead employ two different one-dimensional convolutions for the max-pooled features and average-pooled features, respectively. This approach allows us to store different information and acquire cross-channel interactions without reducing the channel dimensionality. Furthermore, since we use one-dimensional convolution, the increase in the number of parameters is negligible even if no parameters are shared. The specific operation details are outlined below.
%As shown in Fig.~\ref{CAB}, for the given input feature map $\bf{U}\in\mathbb{R}^{H\times W\times C}$, we first exploit both max-pooling and average-pooling operations to gather spatial information. These two simultaneous operations create max-pooled features and average-pooled features respectively: $\bf{U_C^{max}}$and $\bf{U_C^{avg}}$. In SENet \cite{hu2018squeeze}, they designed two fully connected (FC) layers to achieve cross-channel interactions, which brings extra model complexity and the negative effects of channel dimensionality reduction. To overcome this limitation, we introduce two different fast one-dimensional convolutions ($\bf{C1D_k}$) of size $k$ for $\bf{U_C^{max}}$and $\bf{U_C^{avg}}$. The kernel size $k$ stands for the coverage of $k$ neighbors to participate in the interaction between channels. For the definition of $k$, we use the equation in ECA-Net\cite{2020ECA}:
As shown in Fig.~\ref{CAB}, we simultaneously apply max-pooling and average-pooling operations to the input feature map $\bf{U}\in\mathbb{R}^{H\times W\times C}$, generating max-pooled features $\bf{U_C^{max}}$ and average-pooled features $\bf{U_C^{avg}}$, respectively. In contrast to SE \cite{hu2018squeeze}, which used fully connected layers to achieve cross-channel interactions, we use two different one-dimensional convolutions ($\bf{C1D_k}$) of size $k$ for $\bf{U_C^{max}}$ and $\bf{U_C^{avg}}$, respectively, to avoid the negative effects of channel dimensionality reduction and reduce model complexity. The kernel size $k$ is defined as the coverage of $k$ neighbors to participate in the interaction between channels, which is calculated using the equation from ECA-Net \cite{2020ECA}:
\begin{figure}[ht]
\centering
\includegraphics[width=0.48\textwidth]{Images/CAB.pdf}
\caption{Diagram of channel attention block of LHAB. Due to the max-pooling operations and average-pooling operations playing different roles in aggregating spatial dimension information, we design an adaptive channel attention block containing these two operations. The max-pooled and average-pooled features are passed through two separate one-dimensional convolutions, and then activated by the sigmoid function. The resulting vectors are then multiplied by the input feature map for adaptive feature refinement.}
\label{CAB}
\end{figure}
\begin{equation}
k=\psi\left(C\right)=\left| \frac{log_2\left(C\right)}{\gamma}+\frac{b}{\gamma} \right|_{odd}
\end{equation}
where $C$ is the number of channels and $\left| t \right|_{odd}$ represents the nearest odd number of $t$. $\gamma$ and $b$ are set to 2 and 1 respectively in this paper. Through the mapping $\psi$, kernel size $k$ can be adaptively confirmed by the number of channels $C$.
%Then we apply element-wise summation to merge these two feature vectors $\bf{C1D_{k_1}\left(U_C^{max}\right)}$ and $\bf{C1D_{k_2}\left(U_C^{avg}\right)}$, which have gone through different one-dimensional convolutions. The merged result is fed into the sigmoid function and after that, the outcome is multiplied by the original feature map $\bf{U}$ to obtain $\bf{U'}$ for adaptive feature refinement. In a word, the channel attention block is summarized as:
Then we merge these two feature vectors $\bf{C1D_{k_1}\left(U_C^{max}\right)}$ and $\bf{C1D_{k_2}\left(U_C^{avg}\right)}$ using element-wise summation and pass the result through the sigmoid function. The final outcome is obtained by multiplying the original feature map $\bf{U}$ with the result of the sigmoid function to obtain $\bf{U'}$ for adaptive feature refinement. In a word, the channel attention block is summarized as:
\begin{equation}
\begin{split}
\mathbf{U'}&=\sigma \left(\mathbf{C1D_{k_1}}\left(MP\left(\mathbf{U}\right)\right)\oplus \mathbf{C1D_{k_2}}\left(AP\left(\mathbf{U}\right)\right)\right) \otimes \mathbf{U}\\
&=\sigma \left(\bf{C1D_{k_1}\left(U_C^{max}\right)}\oplus \bf{C1D_{k_2}\left(U_C^{avg}\right)}\right) \otimes \bf{U}
\end{split}
\end{equation}
Where $\sigma$ refers to sigmoid function. $MP$ and $AP$ refer to the max-pooling operation and average-pooling operation respectively.
\textbf{Spatial attention block.} To strengthen the inter-spatial relationship of features, we design a spatial attention block. As a complement to channel attention, which pays attention to “what” is essential, spatial attention concentrates on "where" is the important and informative area. Similar to channel attention block, we first apply max-pooling and average-pooling operations along the channel axis to generate two 2D feature maps and then send them to two different two-dimensional convolution layers, which do not share parameters. We describe the detailed operation below.
\begin{figure}[ht]
\centering
\includegraphics[width=0.48\textwidth]{Images/SAB.pdf}
\caption{Diagram of spatial attention block of LHAB. Due to the max-pooling operations and average-pooling operations playing different roles in aggregating channel dimension information, we design an adaptive spatial attention block containing these two operations. Then, the two 2D maps are passed through two different two-dimensional convolutions and further activated by the sigmoid function. Finally, the resulting vectors are multiplied by the input feature map for adaptive feature refinement.}
\label{SAB}
\end{figure}
As shown in Fig.~\ref{SAB}, for the intermediate feature map $\bf{U'}\in\mathbb{R}^{H\times W\times C}$ from the channel attention block, we aggregate channel information by max-pooling and average-pooling operations to obtain two new maps: $\bf{{U'}_S^{max}}\in\mathbb{R}^{H\times W\times 1}$and $\bf{{U'}_S^{avg}}\in\mathbb{R}^{H\times W\times 1}$. Those are then convolved by two different two-dimensional convolution layers ($\bf{C2D_{7\times7}}$), respectively. The kernel size of these two-dimensional convolutions is $7\times7$, which helps to generate larger receptive fields. Then, we merge these two feature maps $\bf{C2D_{7\times7}\left({U'}_S^{max}\right)}$ and $\bf{C2D_{7\times7}\left({U'}_S^{avg}\right)}$ using element-wise summation. The result is activated by the sigmoid function and finally $\bf{U'}$ multiply it to get the end map $\bf{U''}$. In a word, the channel attention block is summarized as:
\begin{equation}
\begin{split}
\mathbf{U''}&=\sigma \left(\mathbf{C2D_{7\times7}}\left(MP\left(\mathbf{U'}\right)\right)\oplus \mathbf{C2D_{7\times7}}\left(AP\left(\mathbf{U'}\right)\right)\right) \otimes \mathbf{U'}\\
&=\sigma \left(\bf{C2D_{7\times7}\left({U'}_S^{max}\right)}\oplus \bf{C2D_{7\times7}\left({U'}_S^{avg}\right)}\right) \otimes \bf{U'}
\end{split}
\end{equation}
where $\sigma$ refers to sigmoid function. $MP$ and $AP$ refer to the max-pooling operation and average-pooling operation respectively.
%In conclusion, in LHAB, channel attention block and spatial attention block are arranged sequentially and channel attention block is in front of spatial attention block. LHAB can make a significant reduction in the number of parameters, the computational effort, and the occupied storage space by the network.
In conclusion, the LHAB module is composed of a channel attention block and a spatial attention block, arranged sequentially with the channel attention block being in front of the spatial attention block. The LHAB can make a significant reduction in the number of parameters, the computational effort, and the occupied storage space of the network, while still effectively capturing important information from the feature maps.
\subsubsection{LHAB-GhostNet}
%The proposed LHAB-GhostNet serves as the backbone for the HSI-ShipDetectionNet, and its structure is outlined in Table~\ref{tab1}. The parameters $Exp$ and $Out$ denote the number of intermediate and output channels, respectively, and $s$ represents the stride. LHAB-GhostNet's architecture is based on that of MobileNetV3\cite{howard2019searching}, with the bottleneck block replaced by our LHAB-Gbneck. The first layer of LHAB-GhostNet is a standard convolution operation, and the network is divided into 5 stages based on input feature map sizes. The stride of the last LHAB-Gbneck in each stage (except for stage 5) is set to 2. Additionally, LHAB is integrated into certain LHAB-Gbnecks, as detailed in Table~\ref{tab1}, to further simplify the backbone and enhance its performance.
The architecture of the proposed LHAB-GhostNet, which serves as the backbone for the HSI-ShipDetectionNet, is summarized in Table~\ref{tab1}. In this table, the parameters $Exp$ and $Out$ indicate the number of intermediate and output channels, respectively, and $s$ represents the stride. The architecture of LHAB-GhostNet is based on MobileNetV3\cite{howard2019searching}, with the bottleneck block replaced by LHAB-Gbneck. The first layer of LHAB-GhostNet is a standard convolution operation, and the network is divided into 5 stages based on the input feature map sizes. The stride of the last LHAB-Gbneck in each stage (except for stage 5) is set to 2. Furthermore, LHAB is integrated into some LHAB-Gbnecks, as illustrated in Table~\ref{tab1}, to further simplify the backbone.
\begin{table}[t!]
\caption{LHAB-GhostNet architecture.}
\label{tab1}
\centering
\renewcommand\arraystretch{1.5}
\begin{tabular}{c|c|c|c|c|c|c}
\toprule
Stage & Input size & Operator & Exp & Out & LHAB & S \\
\midrule
— & $640^2\times3$ & Conv2d & — & 16 & — & 2 \\ \hline
\multirow{2}{*}{stage 1} & $320^2\times16$ & LHAB-Gbneck & 16 & 16 & 0 & 1 \\ \cline{2-7}
& $320^2\times16$ & LHAB-Gbneck & 48 & 24 & 0 & 2 \\ \hline
\multirow{2}{*}{stage 2} & $160^2\times24$ & LHAB-Gbneck & 72 & 24 & 0 & 1 \\ \cline{2-7}
& $160^2\times24$ & LHAB-Gbneck & 72 & 40 & 1 & 2 \\ \hline
\multirow{2}{*}{stage 3} & $80^2\times40$ & LHAB-Gbneck & 120 & 40 & 1 & 1 \\ \cline{2-7}
& $80^2\times40$ & LHAB-Gbneck & 240 & 80 & 0 & 2 \\ \hline
\multirow{6}{*}{stage 4} & $40^2\times80$ & LHAB-Gbneck & 184 & 80 & 0 & 1 \\ \cline{2-7}
& $40^2\times80$ & LHAB-Gbneck & 184 & 80 & 0 & 1 \\ \cline{2-7}
& $40^2\times80$ & LHAB-Gbneck & 184 & 80 & 0 & 1 \\ \cline{2-7}
& $40^2\times80$ & LHAB-Gbneck & 480 & 112 & 1 & 1 \\ \cline{2-7}
& $40^2\times112$ & LHAB-Gbneck & 672 & 112 & 1 & 1 \\ \cline{2-7}
& $40^2\times112$ & LHAB-Gbneck & 672 & 160 & 1 & 2 \\ \hline
\multirow{4}{*}{stage 5} & $20^2\times160$ & LHAB-Gbneck & 960 & 160 & 0 & 1 \\ \cline{2-7}
& $20^2\times160$ & LHAB-Gbneck & 960 & 160 & 1 & 1 \\ \cline{2-7}
& $20^2\times160$ & LHAB-Gbneck & 960 & 160 & 0 & 1 \\ \cline{2-7}
& $20^2\times160$ & LHAB-Gbneck & 960 & 160 & 1 & 1 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{High-Order Spatial Interaction Mechanism}
\label{High-Order Spatial Interaction Mechanism}
%In recent years, the popularity of the Transformer in vision applications has greatly challenged the dominance of CNNs and achieved excellent results in classification, detection, and segmentation tasks. Therefore, some scholars have also applied it to the field of small object detection~\cite{zhu2021tph, gong2022swin}. This is due to Transformer's core architecture, namely Self-Attention, whose success, according to previous research, can be attributed to its ability to capture long-range dependencies, which makes it easier for the model to learn contextual information and facilitates the detection of small objects. Additionally, Self-Attention is able to perform two-order spatial interactions by performing matrix multiplication between queries, keys, and values.
In recent years, the Transformer has gained popularity in vision applications and has challenged the dominance of CNNs by achieving excellent results in classification, detection, and segmentation tasks. Scholars have started exploring the use of Transformer in the field of small object detection, as seen in recent studies such as \cite{zhu2021tph} and \cite{gong2022swin}. The success of Transformer in vision tasks can be attributed to its core architecture, which is self-attention. Self-attention's ability to capture long-range dependencies allows the model to learn contextual information more effectively, which in turn facilitates the detection of small objects. Moreover, self-attention can perform second-order spatial interactions by performing matrix multiplication between queries, keys, and values, which enhances the model's ability to identify spatial relationships.
%However, self-attention has some limitations that need to be addressed. For example, it only stops at two orders for this spatial interaction ability, and Rao et al.\cite{rao2022hornet} demonstrated that higher-order spatial interactions can improve the modeling ability of visual models. Additionally, self-attention requires each token to attend to every other token, which can introduce a quadratic complexity. Finally, it lacks some of the inductive biases compared to CNNs and therefore usually needs to be trained on a considerable amount of data to obtain the desired generalization ability. To overcome these limitations, we introduce a convolution-based architecture, namely Iterative Gated Convolution (g$^n$Conv), to replace self-attention in our method. We take g$^3$Conv as an example to illustrate its principle, as shown in Fig.~\ref{g3Conv}.
Despite its effectiveness, self-attention has some limitations that need to be addressed. For instance, its spatial interaction ability is limited to two orders, while research by Rao et al. \cite{rao2022hornet} has shown that higher-order spatial interactions can improve visual models' modeling ability. Moreover, self-attention introduces a quadratic complexity as it requires each token to attend to every other token. Lastly, self-attention lacks some of the inductive biases present in CNNs, which can make it difficult to generalize well with limited data. To overcome these limitations, we introduce the Iterative Gated Convolution (g$^n$Conv), a convolution-based architecture that replaces self-attention in our method. Specifically, we take g$^3$Conv as an example to illustrate its principle, as shown in Fig.~\ref{g3Conv}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.3\textwidth]{Images/g3Conv.pdf}
\caption{\textbf{g$^3$Conv}. We take g$^3$Conv as an example to illustrate g$^n$Conv's principle. This module can extend the spatial interactions to three orders so that the correlation between features is gradually enhanced through the multiplication.}
\label{g3Conv}
\end{figure}
%For the input feature $x\in\mathbb{R}^{H\times W\times C}$, a linear projection layer (implemented as a convolution operation) is first used to mix the channels. The number of channels is then doubled to obtain the intermediate feature $x'\in\mathbb{R}^{H\times W\times 2C}$. The formula is as follows:
To process the input feature $x\in\mathbb{R}^{H\times W\times C}$, we first use a linear projection layer implemented as a convolution operation to mix the channels. After this operation, the number of channels is doubled to obtain the intermediate feature $x'\in\mathbb{R}^{H\times W\times 2C}$. The formula for this process can be expressed as follows:
\begin{equation}
x'=Conv_{in}\left(x\right)
\end{equation}
Then, the feature map $x'$ is split along the channel dimension, which is expressed as follows:
\begin{equation}
\lbrack a_0, b_0, b_1, b_2\rbrack=Split\left(x'\right)
\end{equation}
where the number of channels for $a_0$ is $\frac{C}{4}$, and the number of channels for $b_0$, $b_1$, and $b_2$ is $\frac{C}{4}$, $\frac{C}{2}$, and $C$, respectively. Then, the depth separable convolution (DW-Conv) is performed on $b_0$, $b_1$, and $b_2$, and the results are iteratively subjected to gated convolution operations with $a_0$, $a_1$, and $a_2$, respectively by:
\begin{equation}
\begin{split}
&a_1=h_0\left(a_0\right)\otimes DW\_Conv_0\left(b_0\right)\\
&a_2=h_1\left(a_1\right)\otimes DW\_Conv_1\left(b_1\right)\\
&a_3=h_2\left(a_2\right)\otimes DW\_Conv_2\left(b_2\right)\\
\end{split}
\end{equation}
where $\otimes$ is the multiplication of the elements in the matrix at the corresponding positions. The role of $\lbrace h_i \rbrace$ is to change the number of channels of $a_i$ to match the number of channels of $b_i$. When $i = 0$, $h_0$ is an identity mapping; when $i$ is 1 or 2, $h_i$ doubles the channels of $a_i$. Finally, the $a_3$ received from the above steps is continued into a linear projection to obtain the final result of g$^3$Conv:
\begin{equation}
y=Conv_{out}\left(a_3\right)
\end{equation}
Based on the above analysis, g$^3$Conv can be generalized to the n-order spatial interaction, i.e. g$^n$Conv. For the input feature map $x\in\mathbb{R}^{H\times W\times C}$, the process is similar to g$^3$Conv, as follows:
\begin{equation}
x'=Conv_{in}\left(x\right)\in\mathbb{R}^{H\times W\times 2C}
\end{equation}
\begin{equation}
\lbrack a_0^{H\times W\times C_0}, b_0^{H\times W\times C_0},\dotsi, b_{n-1}^{H\times W\times C_{n-1}}\rbrack=Split\left(x'\right)
\end{equation}
Where,
\begin{equation}
C_0+\sum_{0\leq i \leq n-1}C_i= 2C
\end{equation}
\begin{equation}
\label{1}
C_i= \frac{C}{2^{n-i-1}}, 0\leq i \leq n-1
\end{equation}
%Equation (\ref{1}) represents how the channel dimension is allocated in each order. This is designed so that fewer channels are used to compute lower orders, thus not introducing a large amount of computational overhead.
Equation (\ref{1}) specifies how the channel dimensions are allocated in each order of the g$^n$Conv operation. This allocation is designed to reduce the number of channels used to compute lower orders, thereby avoiding a large computational overhead. After splitting the intermediate feature map $x'$, the gated convolution continues iteratively:
\begin{equation}
a_{i+1}=h_i\left(a_i\right)\otimes DW\_Conv_i\left(b_i\right), i=0, 1, \dotsi, n-1
\end{equation}
where,
\begin{equation}
\begin{aligned}
h_{i}\left(x\right)= \left \{
\begin{array}{ll}
x, & i=0\\
Conv\left(C_{i-1},C_i\right), & 1\leq i \leq n-1
\end{array}
\right.
\end{aligned}
\end{equation}
The final result for g$^n$Conv is acquired by equation (\ref{2}), as follows:
\begin{equation}
\label{2}
y=Conv_{out}\left(a_n\right)
\end{equation}
%The proposed model, HSI-ShipDetectionNet, utilizes the g$^n$Conv in place of the self-attention mechanism in the Transformer encoder to generate High-Order Spatial Interaction Module (HSI-Former), which is shown in Fig.~\ref{network}. The advantages of g$^n$Conv include the ability to extend spatial interactions to higher orders, which improves feature correlation. Additionally, using a convolution-based architecture avoids the quadratic complexity of self-attention, and the use of channel division further reduces computational cost. Furthermore, convolutional operations introduce inductive biases for ship detection tasks, such as translation equivariance and locality \cite{dosovitskiy2020image}, and this prior knowledge is beneficial for network learning. The depth-separable convolution in the g$^n$Conv uses large $7×7$ convolution kernels to increase the receptive field, which improves context modeling and enhances the detection performance of small ships.
The proposed HSI-ShipDetectionNet model uses g$^n$Conv instead of the self-attention mechanism found in the Transformer encoder to create the High-Order Spatial Interaction (HSI-Former) module. This is illustrated in Figure~\ref{network}. The g$^n$Conv offers several advantages over self-attention, including its ability to extend spatial interactions to higher orders, resulting in improved feature correlation. Moreover, using a convolution-based architecture avoids the quadratic complexity of self-attention, while channel division reduces computational cost. In addition, convolutional operations introduce inductive biases that are helpful for ship detection tasks, such as translation equivariance and locality \cite{dosovitskiy2020image}. This prior knowledge can be beneficial for network learning. In the g$^n$Conv, the depth-separable convolution utilizes large $7×7$ convolution kernels to increase the receptive field. This improves context modeling and enhances the detection performance of small ships.
%\begin{figure}[H]
%\centering
%\includegraphics[width=0.2\textwidth]{Images/HSI_Former.pdf}
%\caption{\textbf{HSI-Former}. The self-attention in the Transformer is replaced with g$^n$Conv and this module is named High-Order Spatial Interaction (HSI-Former) Module }
%\label{HSI-Former}
%\end{figure}
\section{Experiments}
\label{experiments}
\subsection{Experimental Setup}
\subsubsection{Settings}
%All experiments in this paper are performed on a server equipped with NVIDIA Titan V100 GPUs and we implement the deep learning algorithms using PyTorch v1.9.0 and Python v3.8.0. During the training process, the batch size is set to 4. The SGD optimizer is selected for use, in which the momentum and weight decay are set as 0.937 and 5e-4 respectively, and the initial learning rate is 0.01. We stop training after 500 epochs. Rather than searching for the best hyperparameters in the hyperparameter space, we use the same training parameters as those in the corresponding models.
All experiments in this paper are conducted on a server equipped with NVIDIA Titan V100 GPUs, and the deep learning algorithms are implemented using PyTorch v1.9.0 and Python v3.8.0. During the training process, we set the batch size to 4 and use the SGD optimizer with momentum and weight decay of 0.937 and 5e-4, respectively, and an initial learning rate of 0.01. We stop training after 500 epochs. Rather than searching for the best hyperparameters in the hyperparameter space, we use the same training parameters as those in the corresponding models.
\subsubsection{Dataset}
%The data used in the experiments are from a Kaggle competition for marine ship detection\footnote{https://www.kaggle.com/c/airbus-ship-detection}. This dataset contains 29GB of high-resolution optical remote sensing images, with a total of 192,556 images in the training set and 15,606 images in the test set. Each image has a resolution of $768\times768$ pixels. To evaluate the performance of our model in detecting small ships, we select 1000 images from the dataset that contain small target ships and randomly divide them into three subsets: a training set, a validation set, and a test set with a ratio of 7:2:1.
The dataset used in our experiments is sourced from the Kaggle competition for marine ship detection\footnote{https://www.kaggle.com/c/airbus-ship-detection}. The dataset comprises 29GB of high-resolution optical remote sensing images, consisting of a total of 192,556 images in the training set and 15,606 images in the test set. Each image has a resolution of $768\times768$ pixels. To evaluate the effectiveness of our model in detecting small ships, we randomly select 1000 images from the dataset that contain small target ships and divide them into three subsets: a training set, a validation set, and a test set, with a ratio of 7:2:1.
\subsubsection{Evaluation Metrics}
In order to provide a comprehensive evaluation of our proposed method, we consider not only the standard metrics of \textbf{Precision}, \textbf{Recall}, and the \textbf{mean Average Precision (mAP)}, but also the model size, the number of parameters, and the calculated amount. These metrics are commonly used in the field of object detection and can provide a clear understanding of the performance of our model in comparison to other state-of-the-art methods.
These metrics are defined as follows:
\begin{equation}
Recall{=}\frac{TP}{TP+FN}
\end{equation}
\begin{equation}
Precision{=}\frac{TP}{TP+FP}
\end{equation}
\begin{equation}
mAP{=}\int_{0}^{1} Precision \left(Recall\right)d(Recall)
\end{equation}
where TP, FP, and FN represent true positive, false positive, and false negative, respectively, and Precision (Recall) refers to the Precision-Recall curve.
%In addition to this, as one of the focuses of this paper is on the lightweight network, the model size, the number of parameters, and the calculated amount are also considered for evaluation.
\subsection{Comparison with State-of-the-art Methods}
%In this section, our proposed method is compared with a total of three types of models: small object detection models, lightweight detection models, and ship detection models. Finally, we demonstrate the superiority of our method by showing and analyzing the images of the test results.
To evaluate the performance of our proposed method, we compare it with a total of three types of models: small object detection models, lightweight detection models, and ship detection models.
%Note that two models are used to participate in the comparison, namely \textbf{HSI-ShipDetectionNet(V1)} and \textbf{HSI-ShipDetectionNet(V2)}. The original GhostNet is used in HSI-ShipDetectionNet(V1), i.e. the attention mechanism is SE block; HSI-ShipDetectionNet(V2) applies the LHAB-Ghost proposed in this paper.
\subsubsection{Comparison with Small Object Detection Models}
To verify the superior performance of our proposed approach on small object detection, we compare HSI-ShipDetectionNet with two state-of-the-art small object detection models, as described below.
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{TPH-YOLOv5} \cite{zhu2021tph}: This is a YOLOv5-based detector aimed at densely packed small objects. It incorporates advanced techniques such as Transformer blocks, CBAM, and other experienced tricks to improve performance.
\end{itemize}
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{SPH-YOLOv5} \cite{gong2022swin}: The original prediction heads of this detector are replaced with Swin Transformer Prediction Heads (SPHs), which can reduce the computational complexity considerably. In addition, Normalization-based Attention Modules (NAMs) are introduced to improve network detection performance.
\end{itemize}
\begin{table}[ht]
\caption{Comparison of detection performance of different small object detection models}
\label{tab2}
\centering
\setlength{\tabcolsep}{3pt}
\renewcommand\arraystretch{1.5}
\begin{tabular}{c c c c c c}
\hline
\toprule
Models & Para & GFLOPs & R(\%) & mAP(\%) & Size(MB) \\ \toprule
TPH-YOLOv5 & 60.35M & 145.3 & 76.85 & 74.84 & 116.8 \\ \hline
SPH-YOLOv5 & 27.81M & 272.4 & 76.30 & 74.53 & 54.6 \\ \hline
\textbf{HSI-ShipDetectionNet} & \textbf{4.15M} & \textbf{10.0} & 76.85 & 74.35 & \textbf{9.2} \\
\bottomrule
\end{tabular}
\end{table}
%As can be seen in Table~\ref{tab2}, our proposed HSI-ShipDetectionNet has the smallest number of parameters and the smallest amount of computation, so the model takes up the least amount of storage space, only 9.2 MB. Although the mAP value obtained by TPH-YOLOv5 is 0.49 higher than ours, it has 14.5 times the number of parameters and GFLOPs than our model. Similarly, the detection accuracy of SPH-YOLOv5 is comparable to that of HSI-ShipDetectionNet, but the number of parameters and the computational effort of our model are 85.1\% and 96.3\% lower respectively. There is no denying that these two small target detectors have superior detection performance, but they are built on deep and dense convolutional layers. Our proposed model is much lighter than them and achieves a comparable detection accuracy. Considering comprehensively, our method is better suited to scenarios where resources are limited.
As can be seen in Table~\ref{tab2}, our proposed HSI-ShipDetectionNet has the smallest number of parameters and computational complexity, requiring only 9.2 MB of storage space. Although TPH-YOLOv5 achieves a higher mAP value than ours by 0.49, it has 14.5 times more parameters and GFLOPs than our model. Similarly, the detection accuracy of SPH-YOLOv5 is comparable to that of HSI-ShipDetectionNet, but our model requires 85.1\% fewer parameters and 96.3\% less computational effort. While these two small object detectors have superior detection performance, they are built on deep and dense convolutional layers. In contrast, our proposed model is much lighter and achieves comparable detection accuracy. Therefore, our method is better suited for scenarios with limited resources.
\subsubsection{Comparison with Lightweight Detection Models}
To evaluate the performance of our model, we also compare HSI-ShipDetectionNet with the following eight lightweight detection models, described as follows.
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{MobileNetV3-Small} \cite{howard2019searching}: Based on MobileNetV2, MobileNetV3 added the SE block and improved the activation function using h-swish. The small version is targeted at low-resource use cases and therefore contains fewer bottleneck blocks.
\end{itemize}
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{PP-LCNet} \cite{cui2021pp}: This is a lightweight CPU network that utilizes the MKLDNN acceleration strategy. While the techniques used in the network are not novel and have been introduced in previous works, this model achieves a better balance between accuracy and speed through extensive experimentation.
\end{itemize}
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{ShuffleNetV2} \cite{ma2018shufflenet}: Four policies were presented by the authors to reduce memory access costs (MAC), avoid network fragmentation, and reduce element-wise operations.
\end{itemize}
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{MobileNetV3-Large} \cite{howard2019searching}: Unlike MobileNetV3-Small, the large version is targeted at resource-intensive use cases and therefore contains more bottleneck blocks.
\end{itemize}
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{GhostNet} \cite{han2020ghostnet}: It has developed the Ghost module, which tends to accept abundant and redundant information in the feature maps through a cheap operation instead of discarding it.
\end{itemize}
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{Efficient-Lite0} \cite{tan2019efficientnet}: The Efficient-Lite series is the on-device version of EfficientNet and consists of five versions, of which Efficient-Lite0 is the smallest.
\end{itemize}
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{YOLOv5s}: YOLOv5s is the smallest network in the YOLOv5 series in terms of depth and width.
\end{itemize}
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{YOLOv3-tiny} \cite{redmon2018yolov3}: Compared to YOLOv3, YOLOv3-tiny has fewer feature layers and only two prediction branches, making it more suitable for high-speed detection tasks.
\end{itemize}
\begin{table}[ht]
\caption{Comparison of detection performance of different lightweight detection models}
\label{tab3}
\centering
\setlength{\tabcolsep}{3pt}
\renewcommand\arraystretch{1.5}
\begin{tabular}{c c c c c c}
\hline
\toprule
Models & Para & GFLOPs & R(\%) & mAP(\%) & Size(MB) \\ \toprule
MobileNetV3-Small & 3.54M & 6.3 & 71.30 & 68.61 & 7.2 \\ \hline
PP-LCNet & 3.74M & 8.1 & 72.59 & 70.87 & 7.4 \\ \hline
ShuffleNetV2 & 3.78M & 7.7 & 71.30 & 68.86 & 7.5 \\ \hline
MobileNetV3-Large & 5.20M & 10.3 & 72.59 & 70.52 & 10.2 \\ \hline
GhostNet & 5.20M & 8.4 & 73.89 & 71.94 & 10.4 \\ \hline
Efficient-Lite0 & 5.71M & 11.5 & 71.85 & 70.08 & 11.2 \\ \hline
YOLOv5s & 7.05M & 16.3 & 75.92 & 74.25 & 13.7 \\ \hline
YOLOv3-tiny & 8.67M & 12.9 & 75.19 & 73.28 & 16.6 \\ \hline
\textbf{HSI-ShipDetectionNet} & 4.15M & 10.0 & \textbf{76.85} & \textbf{74.35} & 9.2 \\
\bottomrule
\end{tabular}
\end{table}
In order to ensure consistency in experimental conditions, we incorporated the aforementioned lightweight models (excluding YOLOv5s and YOLOv3-tiny) into the framework of YOLOv5 for the purpose of conducting target detection tasks.
As shown in Table~\ref{tab3}, which displays the number of parameters (Para) and recall (R), our proposed HSI-ShipDetectionNet achieves the highest recall and mAP. Compared to the second-best performing model, YOLOv5s, our model not only outperforms in terms of mAP but also has a significantly lower number of parameters, GFLOPs, and model size, at 41.1\%, 38.7\%, and 32.8\% less respectively. Similarly, YOLOv3-tiny has a lower detection accuracy and a more complex network compared to our model. Specifically, our model has a 2.2\% and 1.5\% higher recall rate and mAP respectively, while also having half the number of parameters. This is attributed to the fact that YOLOv3's two prediction branches result in fewer bounding boxes, thus weakening its detection performance. As for GhostNet, it has 5.20M parameters and a 10.4MB model size, 1.05 and 1.2 times higher than our model, but with a 2.41 lower mAP. This demonstrates the superior overall performance of HSI-ShipDetectionNet compared to GhostNet. Additionally, MobileNetV3-Large has a similar GFLOPs as our model, but a 3.83 and 4.26 lower mAP and recall rate respectively. On the other hand, ShuffleNetV2, PP-LCNet and MobileNetV3-Small are indeed lighter than our model, but their detection accuracy (mAP) is around 4 to 6 percentage points lower than that of HSI-SmallShipDetectionNet. These models prioritize lower model complexity over detection accuracy, whereas our HSI-ShipDetectionNet effectively balances both. Overall, HSI-ShipDetectionNet is more sensitive to the detection of small ships while maintaining a suitable level of model complexity.
\subsubsection{Comparison with Ship Detection Models}
To further evaluate the performance of the proposed HSI-ShipDetectionNet in the field of ship detection, we compare it with two state-of-the-art ship detection models. These models are described as follows:
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{ShipDetectionNet} \cite{yin2022enhanced}: This is a lightweight ship detection network that utilizes an improved convolution unit to replace the standard convolution, resulting in a significant reduction in the number of parameters in the network.
\end{itemize}
\begin{itemize}[leftmargin=*,noitemsep]
\item \textbf{Literature} \cite{tan2022improved}: This network proposes a new loss function, IEIOU\_LOSS, and introduces the coordinate attention (CA) mechanism to achieve robust detection results for docked and dense ship targets.
\end{itemize}
\begin{table}[ht]
\caption{Comparison of detection performance of different ship detection models}
\label{tab4}
\centering
\setlength{\tabcolsep}{3pt}
\renewcommand\arraystretch{1.5}
\begin{tabular}{c c c c c c}
\hline
\toprule
Models & Para & GFLOPs & R(\%) & mAP(\%) & Size(MB) \\ \toprule
Literature\cite{tan2022improved} & 7.13M & 16.4 & 73.89 & 71.79 & 14.0 \\ \hline
ShipDetectionNet & 6.05M & 15.7 & 76.11 & 74.26 & 12.0 \\ \hline
\textbf{HSI-ShipDetectionNet} & \textbf{4.15M} & \textbf{10.0} & \textbf{76.85} & \textbf{74.35} & \textbf{9.2} \\
\bottomrule
\end{tabular}
\end{table}
%In the experiments illustrates in Table~\ref{tab4}, we can see that our proposed model performs best in all of the metrics. In terms of mAP, HSI-ShipDetectionNet has almost 3.6\% improvement over the network proposed in the literature\cite{tan2022improved}. However, the number of parameters and the GFLOPs of our model is 41.8\% and 39.0\% lower, respectively, than that of its, so our model consumes less storage space. Compared with ShipDetectionNet, there is a reduction of 31.4\% and 36.3\% regarding parameters and GFLOPs, but our model achieves comparable detection accuracy. This is because the SE attention mechanism is embedded in ShipDetectionNet, and we analyze and improve it by proposing a new Lightweight Hybrid Attention Block (LHAB). It is shown that this new attention mechanism can reduce the computational effort and the number of parameters, and guarantee the detection accuracy of the network. In short, the HSI-ShipDetectionNet is more lightweight and the detection results are more accurate, which is more suitable to ship detection tasks on resource-limited space-borne platforms.
In the experiments illustrated in Table~\ref{tab4}, it can be seen that our proposed model outperforms all the other models in terms of all the evaluation metrics. HSI-ShipDetectionNet has almost 3.6\% higher mAP than the network proposed in the literature \cite{tan2022improved}. Moreover, the number of parameters and GFLOPs of our model is 41.8\% and 39.0\% lower than that of the network in \cite{tan2022improved}, respectively, indicating that our model consumes less storage space. Compared with ShipDetectionNet, our model has a reduction of 31.4\% and 36.3\% regarding parameters and GFLOPs, respectively, while achieving comparable detection accuracy. This is due to the new Lightweight Hybrid Attention Block (LHAB) proposed in our model, which replaces the SE attention mechanism used in ShipDetectionNet. Our analysis shows that the LHAB can reduce the computational effort and the number of parameters while maintaining the detection accuracy of the network. In summary, HSI-ShipDetectionNet is more lightweight and has better detection accuracy, making it more suitable for ship detection tasks on resource-limited space-borne platforms.
\subsubsection{The Visual Comparisons of Different Methods}
%To further demonstrate the superiority of our proposed method for small targets, we show some of the inference results on the test set in Fig.~\ref{results}. It is obvious that HSI-ShipDetectionNet locates and recognizes all small target ships missed by TPH-YOLOv5, GhostNet, and YOLOv5s. Even though ShipDetectionNet also detects all the small ships successfully, the confidence of its prediction box is not as high as that of the HSI-ShipDetectionNet. Meanwhile, our model also shows excellent detection performance for the interference of bad weather such as cloud barriers in the third and seventh columns in Fig.~\ref{results}. Furthermore, while all methods can successfully detect ships in some images, such as the first column of images, the HSI-Former provided by HSI-ShipDetectionNet can more accurately wrap the target ships even though they are difficult to be distinguished with the background by the naked eyes. This is due to the fact that HSI-Former can better understand and model advanced features in deep layers, which is beneficial for improving the accuracy in location and regression for prediction boxes. On the other hand, for some small target ships at the edge of the images, such as those in the second and sixth columns, our proposed model not only identifies them but also detects them with relatively high confidence. This suggests that HSI-ShipDetectionNet can also perform well for more difficult edge targets.
To demonstrate the superior performance of our proposed method for detecting small targets, we present some inference results on the test set in Figure~\ref{results}. It is evident from the results that HSI-ShipDetectionNet successfully locates and recognizes all small target ships that are missed by GhostNet and YOLOv5s. Although ShipDetectionNet also detects all small ships successfully, the confidence of its prediction box is not as high as that of HSI-ShipDetectionNet. In particular, for some images where it is challenging to distinguish the ship from the background, as shown in row (a), HSI-ShipDetectionNet more accurately wraps the target ships. This is due to the fact that HSI-Former can better understand and model advanced features in deep layers, which improves the accuracy of location and regression for prediction boxes. Furthermore, our proposed model can detect small target ships at the edge of the images with relatively high confidence, as shown in rows (c) and (d). Additionally, our model exhibits excellent detection performance in the presence of bad weather conditions, such as cloud barriers shown in row (e). When multiple ships are present in an image, our network can identify all ships more accurately, as shown in row (f).
%In summary, our proposed HSI-ShipDetectionNet improves the detection performance for small-sized ships and is competent with a variety of difficult sea states, giving more reliable and accurate prediction boxes on optical remote sensing images.
To summarize, our proposed HSI-ShipDetectionNet enhances the detection performance of small-sized ships and demonstrates competency in detecting ships in challenging sea conditions. This results in more precise and dependable prediction boxes on optical remote sensing images.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.90\textwidth]{Images/results3.pdf}
\caption{To visualize the inference results from different detection methods on the test set, we will display the outputs of the best-performing model for each method. The methods we comparing are GhostNet, YOLOv5s, ShipDetectionNet, and HSI-ShipDetectionNet.}
\label{results}
\end{figure*}
\subsection{Ablation Experiments and Sensitivity Analysis}
We evaluate the effectiveness of the proposed several modules by ablation analysis and sensitivity analysis.
\subsubsection{Ablation of the Predictive Branch of Tiny Ships}
%To study the influence of the predictive branch of tiny ships ($P_{tiny}$) on detection performance, we first conduct experiments on the detection framework with GhostNet as the backbone. The results of GhostNet on the original detection framework (only three predictive branches) are first obtained experimentally, and then $P_{tiny}$ is added on top of this. The results in Table~\ref{tab5} show that the introduction of the $P_{tiny}$ significantly improves mAP and recall, by 1.07 and 2.22 respectively. This illustrates how the addition of the $P_{tiny}$ branch can improve the network's recall of small ship targets, thus improving detection accuracy.
To study the influence of the predictive branch of tiny ships ($P_{tiny}$) on detection performance, we first conduct experiments on the detection framework with GhostNet as the backbone. We obtain results for GhostNet on the original detection framework (with only three predictive branches), and then add $P_{tiny}$ on top of it. The results in Table~\ref{tab5} show that the introduction of $P_{tiny}$ significantly improves mAP and recall by 1.07 and 2.22, respectively. This indicates that adding the $P_{tiny}$ branch can improve the network's recall of small ship targets, thus improving detection accuracy.
\begin{table}[ht]
\caption{Ablation of the predictive branch of tiny ships}
\label{tab5}
\centering
\setlength{\tabcolsep}{3pt}
\renewcommand\arraystretch{1.5}
\begin{tabular}{c c c c c}
\hline
\toprule
& mAP & Recall(\%) & Parameters & Size(MB)\\ \toprule
GhostNet & 71.94 & 73.89 & 5.20M & 10.4 \\ \hline
\textbf{GhostNet} $\bf{w/ P_{tiny}}$ & \textbf{73.01} & \textbf{76.11} & 5.34M & 11.4 \\
\bottomrule
\end{tabular}
\end{table}
\subsubsection{Ablation and Sensitivity Analysis of the High-Order Spatial Interaction Mechanism}
%Building on the detection framework of the previous part, which has incorporated the $P_{tiny}$ branch, we continue to measure the impact of the High-Order Spatial Interaction (HSI-Former) module on detection performance. The results are shown in Table~\ref{tab6}, where $L$ represents the number of HSI-Former layers and $n$ is the order in g$^n$Conv.
Expanding on the detection framework described in the previous part, which already includes the $P_{tiny}$ branch, we now examine the effects of integrating the High-Order Spatial Interaction (HSI-Former) module on detection performance. Table~\ref{tab6} presents the results of this analysis, where $L$ denotes the number of HSI-Former layers and $n$ refers to the order of g$^n$Conv.
%To investigate the influence of the order, in the case where the number of HSI-Former layers is 1, we vary n from 1 to 4. It is obvious that the model performs the best when the order is 3, and the mAP value is 1.66 higher than that without the HSI-Former module. It also can be seen from them that the worst performance happens when the order is 1. The reason is that 1-order spatial interactions are equivalent to plain convolution, which does not explicitly consider the spatial interactions between a spatial location and its neighboring region\cite{rao2022hornet}. So it contributes little to the performance improvement of the network. In addition, 2-order spatial interactions lead to a slight improvement in the modeling ability of the model, by 0.26. The modeling ability of 4-order spatial interactions is not as good as that of 3-order spatial interactions, with an improvement of only 0.37 compared with the model without the HSI-Former module. This result suggests that it is not that the higher the order of spatial interaction is, the greater the positive impact on the network will be. Further, we also try the effect of 3-order when the HSI-Former layers are 2. It is interesting to see that in the case where layers are 2 when the order of spatial interaction is 3, the model performance is slightly lower than when the HSI-Former layer is 1. This indicates that too many HSI-Former modules may burden the network.
To investigate the impact of the order on model performance, we conduct experiments with varying n from 1 to 4, where the number of HSI-Former layers is fixed at 1. Our findings indicate that the model performs best when the order is 3, with the mAP value 1.66 higher than that without the HSI-Former module. Conversely, the worst performance is observed when the order is 1, as 1-order spatial interactions are equivalent to plain convolution, which fails to explicitly consider spatial interactions between spatial locations and their neighboring regions \cite{rao2022hornet}, thus contributing little to model performance. Furthermore, 2-order spatial interactions show a slight improvement in the modeling ability by 0.26, while 4-order spatial interactions yield an improvement of only 0.37 compared to the model without the HSI-Former module. This result suggests that it is not that the higher the order of spatial interaction is, the greater the positive impact on the network will be. Further, we also try the effect of 3-order when the HSI-Former layers are 2. It is interesting to see that in the case where layers are 2 when the order of spatial interaction is 3, the model performance is slightly lower than when the HSI-Former layer is 1. This indicates that too many HSI-Former modules may burden the network.
%On the other hand, since the HSI-Former is designed on the analysis of the Transformer encoder, we also test the impact of the Transformer block on the network. As can be seen in Table~\ref{tab6}, the size of the model with the Transformer module is comparable to that of the model with HSI-Former(L=1, n=3), but the mAP value drops by 1.04. This strongly suggests that 3-order spatial interactions have more potential for learning and modeling context compared to 2-order spatial interactions.
On the other hand, as the HSI-Former is specifically designed based on the analysis of the Transformer encoder, we conduct a test to evaluate the impact of the Transformer block on the overall network performance. As shown in Table~\ref{tab6}, we observe that the size of the model with the Transformer module is comparable to that of the model with HSI-Former(L=1, n=3). However, the mAP value decreases by 1.04, indicating that 3-order spatial interactions have more potential for learning and modeling context when compared to 2-order spatial interactions. This finding strongly suggests that the HSI-Former architecture with higher order spatial interactions has superior performance in capturing and modeling context for the given task.
\begin{table}[ht]
\caption{Ablation and sensitivity analysis of the high-order spatial interaction mechanism}
\label{tab6}
\centering
\setlength{\tabcolsep}{3pt}
\renewcommand\arraystretch{1.5}
\begin{tabular}{c c c c c}
\hline
\toprule
& mAP & Recall(\%) & Parameters & Size(MB)\\ \toprule
GhostNet w/ $P_{tiny}$ & 73.01 & 76.11 & 5.339M & 11.4 \\ \hline
w/ HSI-Former(L=1, n=1) & 72.70 & 75.19 & 5.631M & 11.9 \\ \hline
w/ HSI-Former(L=1, n=2) & 73.27 & 75.56 & 5.647M & 12.0 \\ \hline
\textbf{w/ HSI-Former(L=1, n=3)} & \textbf{74.67} & \textbf{77.59} & 5.653M & 12.0 \\ \hline
w/ HSI-Former(L=1, n=4) & 73.38 & 76.11 &5.655M & 12.0 \\ \hline
w/ HSI-Former(L=2, n=3) & 73.92 & 76.48 &5.967M & 12.6 \\ \hline
w/ Transformer(L=1) & 73.63 & 76.48 & 5.493M & 11.7 \\
\bottomrule
\end{tabular}
\end{table}
\subsubsection{Ablation of the Lightweight Hybrid Attention Block}
%To further reduce the complexity of the network, we design the Lightweight Hybrid Attention Block (LHAB). The design thinking behind LHAB is expressed in the ablation experiments in this part, the results of which are illustrated in Table~\ref{tab7}. Where ECA(AP) represents the original ECA, that is, there is only an average-pooling operation (AP) operation in ECA. And ECA(MP+AP)$_{share}$ means that both max-pooling operation (MP) and average-pooling operation (AP) are included in the ECA, and that the parameters of both operations are shared. In contrast, $no\_share$ refers to the parameters of the two operations above that are not shared. LHAB adds to ECA(MP+AP)$_{no\_share}$ a spatial attention mechanism that does not share parameters.
To simplify the network further, we design the Lightweight Hybrid Attention Block (LHAB). Our design thinking for LHAB is demonstrated through ablation experiments, and the results are presented in Table~\ref{tab7}. Here, ECA(AP) denotes the original ECA module, where only an average-pooling operation (AP) is employed. ECA(MP+AP)${share}$ implies that both max-pooling operation (MP) and average-pooling operation (AP) are utilized in the ECA module, and the parameters of both operations are shared. Conversely, $no\_share$ indicates that the parameters of these two operations are not shared. LHAB includes a spatial attention mechanism that does not share parameters (Spatial Attention Block) in addition to ECA(MP+AP)${no\_share}$ (Channel Attention Block).
%As shown in Table~\ref{tab7}, the simultaneous existence of both max-pooling operation (MP) and average-pooling operation (AP) improves the feature extraction ability of the network with a 0.13 improvement in mAP over average-pooling operation (AP) only. Furthermore, since max-pooled features and average-pooled features each play different roles, the mAP value is increased by another 0.51 when the parameters of these two operations are not shared. Although $_{no\_share}$ increases the number of parameters in the network, the fact that we use one-dimensional convolutions to extract features means that only 31 (4149741-4149711 = 31) parameters are added, which is completely negligible. ECA(MP+AP)$_{no\_share}$ is the Channel Attention Block we mentioned in Section \ref{method}. Based on this, we proceed to add a Spatial Attention Block that does not share parameters, giving us LHAB. In Table~\ref{tab7}, LHAB achieves the highest mAP (74.35\%) and the number of parameters is reduced by 1.50 (5.65-4.15 = 1.50) million compared to those in Table~\ref{tab6}.
Referring to Table~\ref{tab7}, the inclusion of both max-pooling (MP) and average-pooling (AP) operations in a network enhances its feature extraction ability, resulting in a 0.13 increase in mAP compared to average-pooling operation (AP) alone. Additionally, since max-pooled and average-pooled features have distinct functions, the mAP value is increased by another 0.51 when the parameters of these two operations are not shared. Although this increases the network's parameter count, the use of one-dimensional convolutions for feature extraction means that only 31 (4149741-4149711 = 31) parameters are added, which is insignificant. ECA(MP+AP)${no\_share}$ refers to the Channel Attention Block described in Section \ref{method}. Building upon this, we introduce an independent Spatial Attention Block, which does not share parameters, to create LHAB. In Table~\ref{tab7}, LHAB achieves the highest mAP (74.35\%), while reducing the parameter count by 1.50 (5.65-4.15 = 1.50) million compared to the values presented in Table~\ref{tab6}.
\begin{table}[ht]
\caption{Ablation of the lightweight hybrid attention block}
\label{tab7}
\centering
\setlength{\tabcolsep}{3pt}
\renewcommand\arraystretch{1.5}
\begin{tabular}{c c c c}
\hline
\toprule
& mAP & Recall(\%) & Parameters\\ \toprule
w/o SE w/ ECA(AP) & 73.55 & 75.93 & 4149711 \\ \hline
w/o SE w/ ECA(MP+AP)$_{share}$& 73.68 & 76.48 & 4149711 \\ \hline
w/o SE w/ ECA(MP+AP)$_{no\_share}$ & 74.19 & 76.85 & 4149742 \\
\hline
\textbf{w/o SE w/ LHAB}\\\textbf{(HSI-ShipDetectionNet)} & \textbf{74.35} & \textbf{76.85} & 4150428 \\
\bottomrule
\end{tabular}
\end{table}
\section{Conclusion}
\label{conclusion}
%In this paper, we proposed a novel lightweight ship detection framework specifically designed for small targets. To address the issue of low detection accuracy for small target ships, we introduced a predictive branch for tiny ships, which effectively utilizes the rare pixel information of small targets. Additionally, we proposed a lightweight hybrid attention block (LHAB) to balance detection performance with model complexity by reducing the number of parameters and computational effort. We also incorporated the high-order spatial interaction (HSI-Former) module to enhance the network's ability to understand high-level features, leading to more accurate ship position regression. We conducted comprehensive evaluations of the proposed HSI-ShipDetectionNet, including comparison experiments and ablation studies. The results of these experiments demonstrate the superiority and effectiveness of our proposed framework.
This paper proposes a novel lightweight ship detection framework that is designed specifically for small targets. One of the main challenges with detecting small ships is achieving high detection accuracy due to the scarcity of pixel information. To address this challenge, the proposed framework introduces a predictive branch for tiny ships, which effectively utilizes rare pixel information. In addition, we presents a lightweight hybrid attention block (LHAB) to balance detection performance with model complexity by reducing the number of parameters and computational effort. To enhance the network's ability to understand high-level features, we also incorporates the high-order spatial interaction (HSI-Former) module, which improves the accuracy of ship position regression.
The proposed HSI-ShipDetectionNet is evaluated through comprehensive comparison experiments and ablation studies. The results demonstrate the effectiveness and superiority of the proposed framework in ship detection tasks.
% if have a single appendix:
%\appendix[Proof of the Zonklar Equations]
% or
%\appendix % for no appendix heading
% do not use \section anymore after \appendix, only \section*
% is possibly needed
% use appendices with more than one appendix
% then use \section to start each appendix
% you must declare a \section before using any
% \subsection or using \label (\appendices by itself
% starts a section numbered zero.)
%
% Can use something like this to put references on a page
% by themselves when using endfloat and the captionsoff option.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
% trigger a \newpage just before the given reference
% number - used to balance the columns on the last page
% adjust value as needed - may need to be readjusted if
% the document is modified later
%\IEEEtriggeratref{8}
% The "triggered" command can be changed if desired:
%\IEEEtriggercmd{\enlargethispage{-5in}}
% references section
% can use a bibliography generated by BibTeX as a .bbl file
% BibTeX documentation can be easily obtained at:
% http://mirror.ctan.org/biblio/bibtex/contrib/doc/
% The IEEEtran BibTeX style support page is at:
% http://www.michaelshell.org/tex/ieeetran/bibtex/
%\bibliographystyle{IEEEtran}
% argument is your BibTeX string definitions and bibliography database(s)
%\bibliography{IEEEabrv,../bib/paper}
%
% <OR> manually copy in the resultant .bbl file
% set second argument of \begin to the number of references
% (used to reserve space for the reference number labels box)
\bibliographystyle{IEEEtran}
% argument is your BibTeX string definitions and bibliography database(s)
\normalem
\bibliography{IEEEabrv, ./paper}
% biography section
%
% If you have an EPS/PDF photo (graphicx package needed) extra braces are
% needed around the contents of the optional argument to biography to prevent
% the LaTeX parser from getting confused when it sees the complicated
% \includegraphics command within an optional argument. (You could create
% your own custom macro containing the \includegraphics command to make things
% simpler here.)
%\begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{mshell}}]{Michael Shell}
% or if you just want to reserve a space for a photo:
% insert where needed to balance the two columns on the last page with
% biographies
%\newpage
% You can push biographies down or up by placing
% a \vfill before or after them. The appropriate
% use of \vfill depends on what kind of text is
% on the last page and whether or not the columns
% are being equalized.
%\vfill
% Can be used to pull up biographies so that the bottom of the last one
% is flush with the other column.
%\enlargethispage{-5in}
% that's all folks
\end{document}
% Generated by IEEEtran.bst, version: 1.14 (2015/08/26)
\begin{thebibliography}{10}
\providecommand{\url}[1]{#1}
\csname url@samestyle\endcsname
\providecommand{\newblock}{\relax}
\providecommand{\bibinfo}[2]{#2}
\providecommand{\BIBentrySTDinterwordspacing}{\spaceskip=0pt\relax}
\providecommand{\BIBentryALTinterwordstretchfactor}{4}
\providecommand{\BIBentryALTinterwordspacing}{\spaceskip=\fontdimen2\font plus
\BIBentryALTinterwordstretchfactor\fontdimen3\font minus
\fontdimen4\font\relax}
\providecommand{\BIBforeignlanguage}[2]{{%
\expandafter\ifx\csname l@#1\endcsname\relax
\typeout{** WARNING: IEEEtran.bst: No hyphenation pattern has been}%
\typeout{** loaded for the language `#1'. Using the pattern for}%
\typeout{** the default language instead.}%
\else
\language=\csname l@#1\endcsname
\fi
#2}}
\providecommand{\BIBdecl}{\relax}
\BIBdecl
\bibitem{eldhuset1996automatic}
K.~Eldhuset, ``An automatic ship and ship wake detection system for spaceborne
sar images in coastal regions,'' \emph{IEEE transactions on Geoscience and
Remote Sensing}, vol.~34, no.~4, pp. 1010--1019, 1996.
\bibitem{yin2022enhanced}
Y.~Yin, X.~Cheng, F.~Shi, M.~Zhao, G.~Li, and S.~Chen, ``An enhanced
lightweight convolutional neural network for ship detection in maritime
surveillance system,'' \emph{IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing}, vol.~15, pp. 5811--5825, 2022.
\bibitem{kanjir2018vessel}
U.~Kanjir, H.~Greidanus, and K.~O{\v{s}}tir, ``Vessel detection and
classification from spaceborne optical images: A literature survey,''
\emph{Remote sensing of environment}, vol. 207, pp. 1--26, 2018.
\bibitem{tian2021image}
L.~Tian, Y.~Cao, B.~He, Y.~Zhang, C.~He, and D.~Li, ``Image enhancement driven
by object characteristics and dense feature reuse network for ship target
detection in remote sensing imagery,'' \emph{Remote Sensing}, vol.~13, no.~7,
p. 1327, 2021.
\bibitem{bo2021ship}
L.~Bo, X.~Xiaoyang, W.~Xingxing, and T.~Wenting, ``Ship detection and
classification from optical remote sensing images: A survey,'' \emph{Chinese
Journal of Aeronautics}, vol.~34, no.~3, pp. 145--163, 2021.
\bibitem{pegler2003potential}
K.~H. Pegler, D.~J. Coleman, Y.~Zhang, and R.~P. Pelot, ``The potential for
using very high spatial resolution imagery for marine search and rescue
surveillance,'' \emph{GeoCarto International}, vol.~18, no.~3, pp. 35--39,
2003.
\bibitem{kanjir2014automatic}
U.~Kanjir, A.~Marseti{\v{c}}, P.~Pehani, and K.~O{\v{s}}tir, ``An automatic
procedure for small vessel detection from very-high resolution optical
imagery,'' \emph{Proc. 5th GEOBIA}, pp. 1--4, 2014.
\bibitem{xu2016ship}
F.~Xu and J.-h. Liu, ``Ship detection and extraction using visual saliency and
histogram of oriented gradient,'' \emph{Optoelectronics Letters}, vol.~12,
no.~6, pp. 473--477, 2016.
\bibitem{selvi2011novel}
M.~U. Selvi and S.~S. Kumar, ``A novel approach for ship recognition using
shape and texture,'' \emph{International Journal of Advanced Information
Technology (IJAIT) Vol}, vol.~1, 2011.
\bibitem{xia2011novel}
Y.~Xia, S.~Wan, and L.~Yue, ``A novel algorithm for ship detection based on
dynamic fusion model of multi-feature and support vector machine,'' in
\emph{2011 Sixth International Conference on Image and Graphics}.\hskip 1em
plus 0.5em minus 0.4em\relax IEEE, 2011, pp. 521--526.
\bibitem{tong2020recent}
K.~Tong, Y.~Wu, and F.~Zhou, ``Recent advances in small object detection based
on deep learning: A review,'' \emph{Image and Vision Computing}, vol.~97, p.
103910, 2020.
\bibitem{liu2016ssd}
W.~Liu, D.~Anguelov, D.~Erhan, C.~Szegedy, S.~Reed, C.-Y. Fu, and A.~C. Berg,
``Ssd: Single shot multibox detector,'' in \emph{European conference on
computer vision}.\hskip 1em plus 0.5em minus 0.4em\relax Springer, 2016, pp.
21--37.
\bibitem{lin2017feature}
T.-Y. Lin, P.~Doll{\'a}r, R.~Girshick, K.~He, B.~Hariharan, and S.~Belongie,
``Feature pyramid networks for object detection,'' in \emph{Proceedings of
the IEEE conference on computer vision and pattern recognition}, 2017, pp.
2117--2125.
\bibitem{zagoruyko2016multipath}
S.~Zagoruyko, A.~Lerer, T.-Y. Lin, P.~O. Pinheiro, S.~Gross, S.~Chintala, and
P.~Doll{\'a}r, ``A multipath network for object detection,'' \emph{arXiv
preprint arXiv:1604.02135}, 2016.
\bibitem{guan2018scan}
L.~Guan, Y.~Wu, and J.~Zhao, ``Scan: Semantic context aware network for
accurate small object detection,'' \emph{International Journal of
Computational Intelligence Systems}, vol.~11, no.~1, pp. 951--961, 2018.
\bibitem{ouyang2015deepid}
W.~Ouyang, X.~Wang, X.~Zeng, S.~Qiu, P.~Luo, Y.~Tian, H.~Li, S.~Yang, Z.~Wang,
C.-C. Loy \emph{et~al.}, ``Deepid-net: Deformable deep convolutional neural
networks for object detection,'' in \emph{Proceedings of the IEEE conference
on computer vision and pattern recognition}, 2015, pp. 2403--2412.
\bibitem{zhu2015segdeepm}
Y.~Zhu, R.~Urtasun, R.~Salakhutdinov, and S.~Fidler, ``segdeepm: Exploiting
segmentation and context in deep neural networks for object detection,'' in
\emph{Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition}, 2015, pp. 4703--4711.
\bibitem{zhu2021tph}
X.~Zhu, S.~Lyu, X.~Wang, and Q.~Zhao, ``Tph-yolov5: Improved yolov5 based on
transformer prediction head for object detection on drone-captured
scenarios,'' in \emph{Proceedings of the IEEE/CVF International Conference on
Computer Vision}, 2021, pp. 2778--2788.
\bibitem{chen2021improved}
L.~Chen, W.~Shi, and D.~Deng, ``Improved yolov3 based on attention mechanism
for fast and accurate ship detection in optical remote sensing images,''
\emph{Remote Sensing}, vol.~13, no.~4, p. 660, 2021.
\bibitem{xie2022small}
X.~Xie, L.~Li, Z.~An, G.~Lu, and Z.~Zhou, ``Small ship detection based on
hybrid anchor structure and feature super-resolution,'' \emph{Remote
Sensing}, vol.~14, no.~15, p. 3530, 2022.
\bibitem{xu2022lite}
X.~Xu, X.~Zhang, and T.~Zhang, ``Lite-yolov5: A lightweight deep learning
detector for on-board ship detection in large-scene sentinel-1 sar images,''
\emph{Remote Sensing}, vol.~14, no.~4, p. 1018, 2022.
\bibitem{xu2021board}
P.~Xu, Q.~Li, B.~Zhang, F.~Wu, K.~Zhao, X.~Du, C.~Yang, and R.~Zhong,
``On-board real-time ship detection in hisea-1 sar images based on cfar and
lightweight deep learning,'' \emph{Remote Sensing}, vol.~13, no.~10, p. 1995,
2021.
\bibitem{hu2019sar}
C.~Hu, C.~Chen, C.~He, H.~Pei, and J.~Zhang, ``Sar detection for small target
ship based on deep convolutional neural network,'' \emph{Journal of Chinese
inertial technology}, vol.~27, no.~3, pp. 397--406, 2019.
\bibitem{wu2020coarse}
Y.~Wu, W.~Ma, M.~Gong, Z.~Bai, W.~Zhao, Q.~Guo, X.~Chen, and Q.~Miao, ``A
coarse-to-fine network for ship detection in optical remote sensing images,''
\emph{Remote Sensing}, vol.~12, no.~2, p. 246, 2020.
\bibitem{wang2022yolo}
S.~Wang, S.~Gao, L.~Zhou, R.~Liu, H.~Zhang, J.~Liu, Y.~Jia, and J.~Qian,
``Yolo-sd: Small ship detection in sar images by multi-scale convolution and
feature transformer module,'' \emph{Remote Sensing}, vol.~14, no.~20, p.
5268, 2022.
\bibitem{jin2020patch}
K.~Jin, Y.~Chen, B.~Xu, J.~Yin, X.~Wang, and J.~Yang, ``A patch-to-pixel
convolutional neural network for small ship detection with polsar images,''
\emph{IEEE Transactions on Geoscience and Remote Sensing}, vol.~58, no.~9,
pp. 6623--6638, 2020.
\bibitem{iandola2016squeezenet}
F.~N. Iandola, S.~Han, M.~W. Moskewicz, K.~Ashraf, W.~J. Dally, and K.~Keutzer,
``Squeezenet: Alexnet-level accuracy with 50x fewer parameters and< 0.5 mb
model size,'' \emph{arXiv preprint arXiv:1602.07360}, 2016.
\bibitem{howard2017mobilenets}
A.~G. Howard, M.~Zhu, B.~Chen, D.~Kalenichenko, W.~Wang, T.~Weyand,
M.~Andreetto, and H.~Adam, ``Mobilenets: Efficient convolutional neural
networks for mobile vision applications,'' \emph{arXiv preprint
arXiv:1704.04861}, 2017.
\bibitem{sandler2018mobilenetv2}
M.~Sandler, A.~Howard, M.~Zhu, A.~Zhmoginov, and L.-C. Chen, ``Mobilenetv2:
Inverted residuals and linear bottlenecks,'' in \emph{Proceedings of the IEEE
conference on computer vision and pattern recognition}, 2018, pp. 4510--4520.
\bibitem{howard2019searching}
A.~Howard, M.~Sandler, G.~Chu, L.-C. Chen, B.~Chen, M.~Tan, W.~Wang, Y.~Zhu,
R.~Pang, V.~Vasudevan \emph{et~al.}, ``Searching for mobilenetv3,'' in
\emph{Proceedings of the IEEE/CVF international conference on computer
vision}, 2019, pp. 1314--1324.
\bibitem{zhang2018shufflenet}
X.~Zhang, X.~Zhou, M.~Lin, and J.~Sun, ``Shufflenet: An extremely efficient
convolutional neural network for mobile devices,'' in \emph{Proceedings of
the IEEE conference on computer vision and pattern recognition}, 2018, pp.
6848--6856.
\bibitem{ma2018shufflenet}
N.~Ma, X.~Zhang, H.-T. Zheng, and J.~Sun, ``Shufflenet v2: Practical guidelines
for efficient cnn architecture design,'' in \emph{Proceedings of the European
conference on computer vision (ECCV)}, 2018, pp. 116--131.
\bibitem{han2020ghostnet}
K.~Han, Y.~Wang, Q.~Tian, J.~Guo, C.~Xu, and C.~Xu, ``Ghostnet: More features
from cheap operations,'' in \emph{Proceedings of the IEEE/CVF conference on
computer vision and pattern recognition}, 2020, pp. 1580--1589.
\bibitem{li2020lightweight}
Z.~Li, L.~Zhao, X.~Han, M.~Pan, and F.-J. Hwang, ``Lightweight ship detection
methods based on yolov3 and densenet,'' \emph{Mathematical Problems in
Engineering}, vol. 2020, pp. 1--10, 2020.
\bibitem{jiang2021high}
J.~Jiang, X.~Fu, R.~Qin, X.~Wang, and Z.~Ma, ``High-speed lightweight ship
detection algorithm based on yolo-v4 for three-channels rgb sar image,''
\emph{Remote Sensing}, vol.~13, no.~10, p. 1909, 2021.
\bibitem{liu2022multi}
S.~Liu, W.~Kong, X.~Chen, M.~Xu, M.~Yasir, L.~Zhao, and J.~Li, ``Multi-scale
ship detection algorithm based on a lightweight neural network for spaceborne
sar images,'' \emph{Remote Sensing}, vol.~14, no.~5, p. 1149, 2022.
\bibitem{zheng2022fast}
J.-C. Zheng, S.-D. Sun, and S.-J. Zhao, ``Fast ship detection based on
lightweight yolov5 network,'' \emph{IET Image Processing}, vol.~16, no.~6,
pp. 1585--1593, 2022.
\bibitem{niu2021review}
Z.~Niu, G.~Zhong, and H.~Yu, ``A review on the attention mechanism of deep
learning,'' \emph{Neurocomputing}, vol. 452, pp. 48--62, 2021.
\bibitem{hu2018squeeze}
J.~Hu, L.~Shen, and G.~Sun, ``Squeeze-and-excitation networks,'' in
\emph{Proceedings of the IEEE conference on computer vision and pattern
recognition}, 2018, pp. 7132--7141.
\bibitem{2020ECA}
Q.~Wang, B.~Wu, P.~Zhu, P.~Li, and Q.~Hu, ``Eca-net: Efficient channel
attention for deep convolutional neural networks,'' in \emph{2020 IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR)}, 2020.
\bibitem{huang2019ccnet}
Z.~Huang, X.~Wang, L.~Huang, C.~Huang, Y.~Wei, and W.~Liu, ``Ccnet: Criss-cross
attention for semantic segmentation,'' in \emph{Proceedings of the IEEE/CVF
international conference on computer vision}, 2019, pp. 603--612.
\bibitem{woo2018cbam}
S.~Woo, J.~Park, J.-Y. Lee, and I.~S. Kweon, ``Cbam: Convolutional block
attention module,'' in \emph{Proceedings of the European conference on
computer vision (ECCV)}, 2018, pp. 3--19.
\bibitem{vaswani2017attention}
A.~Vaswani, N.~Shazeer, N.~Parmar, J.~Uszkoreit, L.~Jones, A.~N. Gomez,
{\L}.~Kaiser, and I.~Polosukhin, ``Attention is all you need,''
\emph{Advances in neural information processing systems}, vol.~30, 2017.
\bibitem{dosovitskiy2020image}
A.~Dosovitskiy, L.~Beyer, A.~Kolesnikov, D.~Weissenborn, X.~Zhai,
T.~Unterthiner, M.~Dehghani, M.~Minderer, G.~Heigold, S.~Gelly \emph{et~al.},
``An image is worth 16x16 words: Transformers for image recognition at
scale,'' \emph{arXiv preprint arXiv:2010.11929}, 2020.
\bibitem{rao2022hornet}
Y.~Rao, W.~Zhao, Y.~Tang, J.~Zhou, S.-N. Lim, and J.~Lu, ``Hornet: Efficient
high-order spatial interactions with recursive gated convolutions,''
\emph{arXiv preprint arXiv:2207.14284}, 2022.
\bibitem{redmon2017yolo9000}
J.~Redmon and A.~Farhadi, ``Yolo9000: better, faster, stronger,'' in
\emph{Proceedings of the IEEE conference on computer vision and pattern
recognition}, 2017, pp. 7263--7271.
\bibitem{he2016deep}
K.~He, X.~Zhang, S.~Ren, and J.~Sun, ``Deep residual learning for image
recognition,'' in \emph{Proceedings of the IEEE conference on computer vision
and pattern recognition}, 2016, pp. 770--778.
\bibitem{gong2022swin}
H.~Gong, T.~Mu, Q.~Li, H.~Dai, C.~Li, Z.~He, W.~Wang, F.~Han, A.~Tuniyazi,
H.~Li \emph{et~al.}, ``Swin-transformer-enabled yolov5 with attention
mechanism for small object detection on satellite images,'' \emph{Remote
Sensing}, vol.~14, no.~12, p. 2861, 2022.
\bibitem{cui2021pp}
C.~Cui, T.~Gao, S.~Wei, Y.~Du, R.~Guo, S.~Dong, B.~Lu, Y.~Zhou, X.~Lv, Q.~Liu
\emph{et~al.}, ``Pp-lcnet: A lightweight cpu convolutional neural network,''
\emph{arXiv preprint arXiv:2109.15099}, 2021.
\bibitem{tan2019efficientnet}
M.~Tan and Q.~Le, ``Efficientnet: Rethinking model scaling for convolutional
neural networks,'' in \emph{International conference on machine
learning}.\hskip 1em plus 0.5em minus 0.4em\relax PMLR, 2019, pp. 6105--6114.
\bibitem{redmon2018yolov3}
J.~Redmon and A.~Farhadi, ``Yolov3: An incremental improvement,'' \emph{arXiv
preprint arXiv:1804.02767}, 2018.
\bibitem{tan2022improved}
X.~Tan and Penghui, ``Improve yolov5's ship target detection in sar image,''
\emph{Computer Engineering and Applications}, vol.~58, no.~4, pp. 247--254,
2022.
\end{thebibliography}