Identity-Consistent Multi-Pose Generation of Contactless Fingerprints
Extracted source text ยท arXiv 2605.03830
%% Identity-Consistent Multi-Pose Generation of Contactless Fingerprints (IMPOSE)
%% For submission to IEEE Transactions on Information Forensics and Security (TIFS)
\documentclass[journal]{IEEEtran}
% *** PACKAGES ***
\usepackage{graphicx}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{multirow}
\usepackage{booktabs}
\usepackage{threeparttable}
\usepackage{makecell}
\usepackage[table]{xcolor}
\usepackage{subfig}
\usepackage{url}
\usepackage{array}
\usepackage{orcidlink}
\usepackage{microtype}
\widowpenalty=10000
\clubpenalty=10000
\hyphenation{}
\hypersetup{hidelinks}
\begin{document}
\title{Identity-Consistent Multi-Pose Generation of Contactless Fingerprints}
\author{Zhiyu Pan$^{\orcidlink{0009-0000-6721-4482}}$,
Xiongjun Guan$^{\orcidlink{0000-0001-8887-3735}}$,
Jianjiang Feng$^{\orcidlink{0000-0003-4971-6707}}$, ~\IEEEmembership{Member, IEEE},
and Jie Zhou$^{\orcidlink{0000-0001-7701-234X}}$, ~\IEEEmembership{Fellow, IEEE}
\thanks{
This work was supported in part by the National Natural Science Foundation of China under Grant 62376132 and 62321005. (\emph{Corresponding author: Jianjiang Feng}.)}
\IEEEcompsocitemizethanks{
\IEEEcompsocthanksitem
The authors are with Department of Automation, Tsinghua University, Beijing 100084, China (e-mail: \url{pzy20@mails.tsinghua.edu.cn}; \url{gxj21@mails.tsinghua.edu.cn};
\url{jfeng@tsinghua.edu.cn}; \url{jzhou@tsinghua.edu.cn}).}}
\markboth{IEEE Transactions on Information Forensics and Security}%
{Pan \MakeLowercase{\textit{et al.}}: Identity-Consistent Multi-Pose Generation of Contactless Fingerprints}
\maketitle
\begin{abstract}
Contactless fingerprint recognition has gained increasing attention due to its advantages in hygiene, acquisition flexibility, and device compatibility. However, the absence of physical contact constraints introduces severe nonlinear geometric distortions caused by free finger poses in 3D space, resulting in a substantial cross-modal domain gap between contactless and conventional contact-based fingerprints. Existing solutions largely rely on explicit geometric correction or image enhancement as preprocessing, which are fragile under extreme pose variations. In this paper, we approach the problem from a data-centric perspective and propose Identity-Consistent Multi-Pose Generation of Contactless Fingerprints (IMPOSE), a physics-inspired framework that synthesizes identity-preserving, multi-pose contactless fingerprint samples to empower recognition models. IMPOSE consists of three core stages: (1) rolled fingerprint identity generation via latent diffusion with discrete codebook representations, (2) cross-modal translation from rolled to contactless modality guided by Sauvola-based local adaptive binarization as an identity anchor, and (3) physics-based multi-pose simulation through 3D finger model texture mapping and projection. The generated samples maintain strict identity consistency at the ridge topology level and spatial alignment with standard fingerprint coordinate space. Extensive experiments on the UWA and PolyU CL2CB databases demonstrate that fine-tuning fixed-length dense descriptors (FDD) with IMPOSE-synthesized data achieves state-of-the-art cross-modal matching performance, reducing EER to 8.74\% on UWA (all poses) and 2.26\% on PolyU CL2CB. Moreover, the synthetic data yields consistent performance gains across mainstream fixed-length representations including DeepPrint and AFRNet, and the hybrid fine-tuning strategy combining synthetic and real data achieves the best overall results, demonstrating the strong substitution potential of high-quality physically-aligned synthetic data. The code and generated samples will be made publicly available at \url{https://github.com/Yu-Yy/IMPOSE}.
\end{abstract}
\begin{IEEEkeywords}
Contactless fingerprint, fingerprint generation, diffusion model, cross-modal matching, multi-pose simulation, fixed-length representation.
\end{IEEEkeywords}
\IEEEpeerreviewmaketitle
\section{Introduction}
\IEEEPARstart{C}{ontactless} fingerprint recognition has emerged as a prominent research direction owing to its superior hygiene, user convenience, and compatibility with ubiquitous camera-equipped mobile devices~\cite{grosz2021c2cl}. Unlike contact-based acquisition that constrains the finger on a flat platen, contactless capture allows free finger placement in 3D space, naturally eliminating cross-infection risks and broadening application scenarios. Compared to face biometrics, fingerprints offer stronger privacy protection and resistance to spoofing, as they are covert traits not readily exposed in public~\cite{maltoni2022handbook}. Despite these advantages, the freedom in finger placement introduces severe multi-pose challenges: rotation, pitch, and yaw of the finger in 3D space produce significant nonlinear geometric distortions and occlusions, creating a substantial cross-modal domain gap between contactless images and conventional contact-based gallery fingerprints.
To address these challenges, mainstream research paradigms have largely focused on image enhancement~\cite{grosz2021c2cl}, cross-modal spatial alignment~\cite{dong2023synthesis,dong2025bridging}, and geometric distortion correction~\cite{cui2023monocular} as preprocessing steps. These methods aim to narrow the domain gap by standardizing contactless images to resemble contact-based counterparts, thereby enabling the extraction of unified discriminative features. However, the inherent complexity of nonlinear finger deformations and the fragility of explicit geometric reconstruction often limit their robustness in extreme pose scenarios.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{figures/impose_motivation_en.pdf}
\caption{Comparison of strategies for improving contactless fingerprint matching performance. (a) The conventional paradigm relies on explicit geometric correction and enhancement. (b) Our IMPOSE framework addresses the problem from a data-centric perspective, empowering recognition models through identity-consistent synthetic data.}
\label{fig:motivation}
\end{figure}
In prior work~\cite{pan2024FDD,pan2025FLARE}, a Fixed-length Dense Descriptor (FDD) was developed that demonstrates strong intrinsic representation robustness in global perception. Remarkably, even when trained solely on high-quality rolled fingerprint data, FDD achieved competitive matching performance on multiple heterogeneous fingerprint databases, including PolyU CL2CB~\cite{liu20143d}. This key observation motivates a fundamentally different technical route: rather than designing fragile and complex explicit image reconstruction pipelines, we propose to empower the representation network through identity-consistent, multi-pose synthetic samples that cover diverse acquisition conditions. By leveraging physics-simulated contactless multi-modal samples for end-to-end learning, the recognition system can implicitly learn robust features resilient to nonlinear distortions and illumination variations, without relying on cumbersome preprocessing operators (see Fig.~\ref{fig:motivation}).
In fingerprint generation, deep learning has driven a paradigm shift from hand-crafted statistical modeling~\cite{cappelli2004sfinge} toward data-driven synthesis~\cite{engelsma2022printsgan,grosz2024universal,shoshan2024fpgan}. Recent advances such as GenPrint~\cite{grosz2024universal}, built on denoising diffusion probabilistic models (DDPM)~\cite{ho2020DDPM}, can simulate fingerprints under various sensor styles and text-driven conditions. However, existing methods predominantly focus on conventional contact-based planar or rolled fingerprints, with limited exploration of contactless modalities. Notable exceptions like SynColFinger~\cite{priesnitz2022syncolfinger} rely on traditional SFinGe~\cite{cappelli2004sfinge} texture models, resulting in significant discrepancies in ridge detail and illumination realism. While general-purpose models such as GenPrint can mimic contactless styles, they operate as style transfer on planar fingerprints and cannot physically simulate the nonlinear distortions and multi-pose geometric transformations inherent to unconstrained contactless acquisition.
Motivated by the above observations, we propose Identity-Consistent Multi-Pose Generation of Contactless Fingerprints (IMPOSE), a complete physics-inspired generation pipeline. IMPOSE first synthesizes rolled fingerprints as original identity references, then translates them to the contactless domain through a cross-modal conversion mechanism, and finally simulates diverse acquisition poses and nonlinear distortions via 3D finger point cloud projection. For identity consistency control, we introduce the Sauvola local adaptive thresholding algorithm~\cite{sauvola2000adaptive} as a cross-modal condition. Unlike deep learning-based binarization methods that may produce hallucinated textures, Sauvola is a deterministic digital image processing technique that stably extracts ridge structures in a non-learning manner, ensuring the fidelity of identity information during cross-modal generation.
Using large-scale datasets generated by IMPOSE, we fine-tune several representative fixed-length fingerprint descriptor extraction networks to enhance their matching robustness in contactless scenarios. Extensive quantitative and qualitative experiments validate the richness of generated samples and their ridge-level identity consistency. Verification on the multi-pose contactless database UWA~\cite{zhou2014benchmark} and the cross-modal database PolyU CL2CB~\cite{liu20143d} demonstrates that synthetic data significantly boosts the performance of fixed-length representations, offering a novel data-driven perspective for solving the contactless fingerprint recognition challenge.
The main contributions of this work are:
\begin{enumerate}
\item We propose IMPOSE, a data-centric framework that approaches contactless fingerprint matching from a generation-and-augmentation perspective. By synthesizing identity-consistent, multi-pose contactless samples for augmentation, IMPOSE enables fixed-length representations to learn robustness against nonlinear distortions, avoiding the fragility of explicit geometric correction.
\item To preserve identity fidelity during cross-modal generation, we introduce Sauvola local adaptive binarization as a deterministic, non-learned identity anchor that avoids hallucination artifacts. For realistic multi-pose distortion, we construct a physics-based simulation pipeline using 3D finger model texture mapping and projection, ensuring synthetic samples faithfully reflect the nonlinear geometric variations encountered in real contactless acquisition.
\item Extensive experiments on UWA and PolyU CL2CB databases demonstrate that IMPOSE-synthesized data consistently enhances mainstream fixed-length representations (FDD, DeepPrint, AFRNet), achieving state-of-the-art cross-modal matching performance. Systematic comparison with real data fine-tuning further validates the strong substitution potential of high-quality physically-aligned synthetic data.
\end{enumerate}
\section{Related Work}
\begin{figure*}[!t]
\centering
\includegraphics[width=.95\textwidth]{figures/impose_pipeline.pdf}
\caption{Overview of the IMPOSE framework for multi-pose contactless fingerprint generation.}
\label{fig:pipeline}
\end{figure*}
\subsection{Deep Generative Models for Fingerprint Synthesis}
Fingerprint generation has evolved from physics-based manual modeling to data-driven deep generative approaches. The seminal SFinGe~\cite{cappelli2004sfinge} established a complete physical simulation framework using mathematical ridge flow models with Gabor filters, enabling the generation of multiple impressions from the same virtual finger. Subsequent works introduced statistical modeling of ridge features~\cite{zhao2012fingerprint} and minutiae-guided generation~\cite{johnson2013texture}, but remained constrained by their hand-crafted nature, producing relatively homogeneous patterns lacking complex intra-class variations.
With the rise of deep learning, Generative Adversarial Networks (GANs)~\cite{goodfellow2020GAN} became the dominant paradigm for fingerprint synthesis. Early GAN-based methods~\cite{mistry2020fingerprint,minaee2018finger,riazi2020synfi,bahmani2021high,cao2018fingerprint} significantly improved texture realism but lacked explicit control over fingerprint identity. PrintsGAN~\cite{engelsma2022printsgan} marked a milestone by introducing a two-stage generation paradigm: it first generates a binary ``Master Print'' as identity reference, then simulates different impressions through latent-variable-driven deformation. This decoupling of identity and appearance inspired subsequent work. FPGAN-Control~\cite{shoshan2024fpgan} combined VAE and GAN architectures for fine-grained appearance control through latent space interpolation.
More recently, Denoising Diffusion Probabilistic Models (DDPM)~\cite{ho2020DDPM} have opened new avenues for controllable fingerprint generation. GenPrint~\cite{grosz2024universal} exploited DDPM's text-guided control capabilities to generate fingerprints conditioned on human-interpretable attributes including pattern type, sensor type, and quality level, while maintaining identity consistency through a pre-trained ridge extractor.
Despite these advances, existing methods predominantly target contact-based fingerprints. For contactless modalities, GenPrint~\cite{grosz2024universal} provides preliminary contactless-style generation but operates as style transfer on planar fingerprints without modeling 3D pose variations. SynColFinger~\cite{priesnitz2022syncolfinger} overlays contactless color statistics on SFinGe-generated ridges with hand-crafted distortion fields, resulting in low texture fidelity. Dong et al.~\cite{dong2023synthesis} explored multi-view 3D fingerprint simulation by mapping SFinGe textures onto Bezier-fitted 3D finger surfaces, but the generated textures remain overly idealized. Building a framework that achieves both high-fidelity texture generation and physically-realistic multi-pose distortion simulation remains an open challenge.
\subsection{Contactless Fingerprint Matching and Recognition}
Contactless fingerprint matching has been approached from two main angles: local feature modeling and geometric distortion correction. For local feature modeling, Kumar et al.~\cite{kumar2013towards} proposed 3D finger surface codes extending minutiae into 3D space. Tan et al.~\cite{tan2020towards} employed elliptical finger models with singular point information for pose estimation and improved minutiae extraction. Shi et al.~\cite{shi2022towards} explored CNN- and GCN-based architectures for robust 2D minutiae representation. Dong et al.~\cite{dong2025bridging} achieved robust recovery of 3D minutiae from single 2D contactless images, bridging the modality gap between contactless 3D features and contact-based 2D templates.
For geometric correction, Cui et al.~\cite{cui2023monocular} and Jia et al.~\cite{jia2024improving} predicted finger depth from monocular images and performed unwarping before feature matching. Grosz et al.~\cite{grosz2021c2cl} proposed C2CL, which adaptively learns elastic distortion parameters guided by fixed-length representation similarity, achieving strong cross-modal performance under moderate pose variations.
In contrast to these explicit correction approaches, we argue that with sufficiently powerful deep representations and large-scale, spatially-diverse training data, models can implicitly learn to handle nonlinear distortions without fragile explicit geometric preprocessing. This data-driven philosophy motivates our IMPOSE framework.
\section{Methodology}
The IMPOSE framework, illustrated in Fig.~\ref{fig:pipeline}, consists of three core stages: (1) rolled fingerprint identity generation, which establishes unique fingerprint identities as the foundational reference; (2) cross-modal translation from rolled to contactless modality guided by Sauvola binary enhancement maps as identity anchors; and (3) physics-based multi-pose simulation through 3D finger model texture mapping and projection. The following subsections detail each stage.
\subsection{Preliminaries: Denoising Diffusion Probabilistic Models}
Denoising Diffusion Probabilistic Models (DDPM)~\cite{ho2020DDPM} are a class of Markov chain-based generative models that transform complex data distributions into simple isotropic Gaussian distributions through simulated diffusion processes.
DDPM consists of a forward diffusion process and a reverse denoising process. The forward process defines a Markov chain by gradually adding Gaussian noise $\epsilon$ to the original data $x_0$ over $T$ timesteps. Using the reparameterization trick, the noisy sample $x_t$ at any timestep $t$ can be directly expressed as:
\begin{equation}
x_t = \sqrt{\bar{\alpha}_t}x_0 + \sqrt{1 - \bar{\alpha}_t}\epsilon, \quad \epsilon \sim \mathcal{N}(\mathbf{0}, \mathbf{I})
\label{eq:ddpm_add}
\end{equation}
where $\alpha_t = 1 - \beta_t$ is a predefined hyperparameter and $\bar{\alpha}_t := \prod_{s=0}^{t} \alpha_s$ denotes the cumulative noise coefficient. The training objective is to learn a parameterized network $\epsilon_\theta$ that predicts the noise component added to the data.
To improve computational efficiency for high-resolution image generation, we adopt the Latent Diffusion Model (LDM)~\cite{rombach2022LDM} paradigm, which performs the diffusion process in a low-dimensional latent space constructed by a pretrained variational autoencoder. Under this framework, the encoder $\xi$ maps an input image $x_i$ to a latent representation $z_0 = \xi(x_i)$, allowing the denoising network $\epsilon_\theta$ to operate in a computationally efficient latent space. With cross-modal conditioning constraints (e.g., identity features), the training objective is:
\begin{equation}
\mathcal{L} = \mathbb{E}_{z_t, t, C, \epsilon \sim \mathcal{N}(0,1)} \left[ \left\| \epsilon - \epsilon_\theta(z_t, t, C) \right\|^2_2 \right]
\label{eq:ddpm_loss}
\end{equation}
where $C$ represents the conditional embedding extracted by a pretrained encoder. By injecting this conditional signal during denoising, the model can precisely control both global structural properties and fine-grained texture details of the generated images.
During inference, we adopt the DDIM~\cite{song2020denoising} formulation for iterative denoising from random Gaussian noise $z_T$ to the final latent representation $z_0$. DDIM enables significantly fewer inference steps than training without sacrificing quality. By setting the stochastic variance term $\sigma_t$ to 0, the reverse process becomes a deterministic mapping:
\begin{equation}
z_{t-1} = \sqrt{\bar{\alpha}_{t-1}} \hat{z}_0 + \sqrt{1 - \bar{\alpha}_{t-1}} \cdot \epsilon_\theta(z_t, t, C),
\label{eq:ddim}
\end{equation}
where $\hat{z}_0 = (z_t - \sqrt{1 - \bar{\alpha}_t} \epsilon_\theta(z_t, t, C)) / \sqrt{\bar{\alpha}_t}$ is the predicted clean latent at timestep $t$.
This deterministic sampling trajectory enables accelerated inference while ensuring stable identity information transmission throughout the generation chain.
\subsection{Rolled Fingerprint Identity Generation}
The first stage of IMPOSE establishes unique fingerprint identities as the foundational reference for the entire pipeline. We employ a hierarchical approach combining discrete latent space modeling with latent diffusion.
For image compression, we introduce a Vector Quantized Variational Autoencoder (VQ-VAE)~\cite{vqvae} to discretize the fingerprint latent space, constructing a structured codebook. By mapping complex local texture features to finite codebook indices, VQ-VAE achieves precise feature quantization and effective redundancy reduction. This discrete representation captures global topological patterns through modeling the joint distribution of code indices, accurately preserving ridge flow continuity and minutiae uniqueness.
The encoder $E_I$ first maps an input rolled fingerprint image $I$ to continuous latent representations $z_l$. Vector quantization against codebook yields discrete latent representations $z_i$, which are then decoded by $D_I$ to produce the rolled fingerprint image $I_r$. Considering the complexity of ridge structures, we employ only $4\times$ spatial downsampling with 3 latent channels to prevent irreversible loss of critical spatial information. The codebook size is $N_C = 4,096$ with each feature vector of dimension 3.
In the diffusion stage, we train a U-Net-based denoising network on the codebook-indexed latent space to synthesize novel fingerprint identities from Gaussian noise. At inference, the system samples random noise $z_T$ in the latent space and iteratively denoises it through $T$ timesteps. To ensure the generated latent vectors respect fingerprint manifold constraints, the derived $\hat{z}_0$ is further refined through codebook-based vector quantization before being decoded to pixel space by the pretrained decoder $D_I$ (see Fig.~\ref{fig:rolled_gen_pipeline}).
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{figures/Rolled_Gen_Network.pdf}
\caption{Process of rolled fingerprint synthesis with corresponding generated examples.}
\label{fig:rolled_gen_pipeline}
\end{figure}
The training data consists of 1,093 high-resolution optical rolled fingerprints collected in-house (THU Rolled) and 2,000 rolled fingerprint samples from NIST SD302~\cite{nist302}, totaling 3,093 images. All samples are preprocessed with high-precision pose estimation~\cite{duan2023estimating} for alignment and normalized to $512 \times 512$ pixels at 500 PPI.
\subsection{Identity-Consistent Control via Sauvola Binarization}
\label{sec:sauvola}
The core challenge in rolled-to-contactless modality translation lies in balancing cross-modal style realism with identity invariance. Unlike face recognition focused on global semantics, fingerprint recognition demands extremely low tolerance for ridge and minutiae deviations---subtle alterations in ridge topology can lead to identity shift.
Through empirical investigation, we found that conventional approaches such as fixed-length descriptor supervision~\cite{pan2024FDD}, ridge extractor-based supervision~\cite{nist2020verifinger}, or minutiae-based rigid alignment of training pairs all struggle to achieve effective identity consistency control during generation training, primarily due to inherent differences in acquisition time, contact area, and nonlinear distortion between rolled and contactless fingerprints.
We propose a novel control perspective: achieving ridge-level pixel correspondence between the two modalities naturally ensures precise identity preservation without additional loss functions. Drawing inspiration from prior work~\cite{engelsma2022printsgan,grosz2024universal} that uses binary ridge enhancement maps as identity injection signals, we identify two key mechanisms: (1) binary images provide high contrast that clearly characterizes identity features; (2) binary maps extracted directly from the target modality ensure natural ridge-level consistency.
However, contactless fingerprints typically exhibit low contrast, causing mature ridge extraction methods to either fail or produce hallucinated textures that induce identity drift (see the comparison in Fig.~\ref{fig:enh_samples}). To address this, we adopt the Sauvola local adaptive binarization algorithm~\cite{sauvola2000adaptive}, a deterministic image processing technique that adaptively estimates local statistical properties for each pixel neighborhood.
For a fingerprint grayscale image $I(x,y)$, the local mean $m(x,y)$ and standard deviation $s(x,y)$ are computed within a $w \times w$ window:
\begin{equation}
\begin{aligned}
m(x,y) &= \frac{1}{w^2} \sum_{i=x-r}^{x+r} \sum_{j=y-r}^{y+r} I'(i,j), \\
s(x,y) &= \sqrt{\frac{1}{w^2} \sum_{i=x-r}^{x+r} \sum_{j=y-r}^{y+r} [I'(i,j) - m(x,y)]^2},
\end{aligned}
\end{equation}
where $r = (w-1)/2$ and $w = 11$ to match typical fingerprint ridge width. The local threshold is then:
\begin{equation}
T(x,y) = m(x,y) \left[ 1 + k \left( \frac{s(x,y)}{R} - 1 \right) \right],
\end{equation}
where $R = 128$ is the dynamic range constant and $k = 0.007$ is the correction parameter. This mechanism automatically adapts discrimination criteria between ridge regions (high $s$) and background regions (low $s$), suppressing noise while enhancing valid texture.
The final binary image is obtained by:
\begin{equation}
B(x,y) = \begin{cases} 0, & \text{if } (x,y) \in \mathcal{M} \text{ and } I'(x,y) > T(x,y) \\ 255, & \text{otherwise} \end{cases}
\end{equation}
followed by morphological opening with a $2 \times 2$ structuring element to remove residual noise.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{figures/enh_samples.pdf}
\caption{Comparison of Sauvola binarization against VeriFinger v12.0 on contactless fingerprints.}
\label{fig:enh_samples}
\end{figure}
Although Sauvola-generated ridge edges are slightly less smooth than those from specialized extractors, its deterministic nature maximizes fidelity to the original texture, avoiding the ``blind guessing'' common in learned methods on feature-sparse regions. This aligns with the enhancement principle established in prior work~\cite{pan2024FDD,pan2025FLARE} of avoiding explicit ridge perception logic in favor of deterministic texture mapping.
\subsection{Contactless Fingerprint Modality Translation}
\label{sec:cl_gen}
Building on the Sauvola-derived image $B$, we construct a high-fidelity cross-modal translation pipeline. Following GenPrint's identity injection logic~\cite{grosz2024universal}, we combine identity binary image latent encoding with ControlNet~\cite{zhang2023adding} injection.
The translation module training is organized into three sequential stages:
\textbf{Stage 1: Binary Image Latent Space Modeling.} To effectively inject the Sauvola binary image $B$ into the diffusion process, we train a dedicated autoencoder ($E_{\text{e}}$, $D_{\text{e}}$) that maps binary images to a latent representation $z_B$ with dimensions matching the denoising latent space. The architecture mirrors VQ-VAE but uses continuous latent vectors without quantization to preserve richer detail. Training employs MSE loss to ensure $z_B$ faithfully reconstructs the ridge topology.
\textbf{Stage 2: Contactless Modality Prior Learning.} A VQ-VAE ($E_{\text{cl}}$, $D_{\text{cl}}$) establishes the compressed representation space for contactless modality, followed by LDM training to capture modality-specific texture statistics. This pretrained backbone ensures generated images strictly conform to contactless acquisition norms.
\textbf{Stage 3: Identity-Consistent Translation.} We introduce the ControlNet architecture~\cite{zhang2023adding} with parameter-decoupled training: only the ControlNet side-branch weights are updated while the backbone denoising network remains frozen. ControlNet's encoder is initialized from the pretrained contactless generation module, with zero convolution layers embedded after each convolution block to prevent disruptive noise during early training stages. This design establishes a closed-loop control framework that achieves pixel-level identity transfer from $I_r$ to $I_cl$ solely through spatial guidance from Sauvola binary maps, without requiring additional identity loss supervision.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{figures/CB2CL.pdf}
\caption{Training stages and architecture of the contactless fingerprint cross-modal translation module.}
\label{fig:CB2CL}
\end{figure}
For training, we utilize the UWA~\cite{zhou2014benchmark} contactless fingerprint database, which provides flattened contactless images with excellent field-of-view coverage suitable for subsequent multi-pose texture mapping. All contactless images are normalized to $512 \times 512$ pixels at 500 PPI, ensuring spatial alignment with the rolled fingerprint generation module.
\subsection{Multi-Pose Contactless Fingerprint Projection Simulation}
\label{sec:multi_pose}
\begin{figure*}[!t]
\centering
\includegraphics[width=\textwidth]{figures/multi-pose-pipeline.png}
\caption{Illustration of the multi-pose contactless fingerprint simulation pipeline: from 3D model unfolding, pose estimation, and texture mapping to final multi-angle projection.}
\label{fig:multi_pose_pipeline}
\end{figure*}
To evolve synthesized contactless textures under diverse 3D poses, we construct a simulation pipeline using finger surface point clouds as geometric carriers. The core logic establishes a mapping between 2D textures and 3D geometric manifolds to enable multi-angle projection.
We first employ the parametric unfolding algorithm proposed by Guan et al.~\cite{guan20213dpose} to map self-collected 3D fingerprint point clouds (THU Finger3D) to a 2D UV plane. Pose estimation~\cite{duan2023estimating} is then applied to rectify the unfolded fingerprint image, aligning the UV space with the standard fingerprint coordinate system. This coordinate alignment ensures the generated contactless texture can establish valid point-to-point correspondences with the 3D point cloud.
By precisely injecting discrete texture color information into the finger geometric model, the system can stably generate contactless fingerprint images under various viewing poses by altering the observation angle in 3D space.
Our multi-pose simulation focuses on the finger roll angle, which is the primary challenge in contactless fingerprint recognition as it directly governs nonlinear ridge distortion~\cite{zhou2014benchmark, dong2023synthesis, tan2020towards, cui2023monocular}. We model finger rolling as a rigid-body pure rolling process on the UV plane, establishing a dynamic correspondence between 3D Euclidean space and 2D UV space, where the UV space coincides with the standard fingerprint coordinate system (see Fig.~\ref{fig:uv23d}).
\begin{figure}[!t]
\centering
\subfloat[Front view]{\includegraphics[height=0.3\linewidth]{figures/front_view_finger.png}} \hfil
\subfloat[Top view]{\includegraphics[height=0.3\linewidth]{figures/top_view_finger.png}}
\caption{Illustration of the finger roll projection mechanism under different geometric views.}
\label{fig:uv23d}
\end{figure}
During 3D model unfolding, after initial pose correction, point clouds are segmented along the principal $y$-axis with the model centroid as origin. Geodesic distances from each sampling point to the $x=0$ reference plane are computed and projected onto the $u$ coordinate axis of the 2D parameter space. Under this modeling, the $u$-$v$ axes of the UV plane maintain parallel correspondence with the $x$-$y$ axes of 3D space.
However, directly orthographically projecting the rotated 3D model onto the $x$-$y$ plane would deviate from the standard fingerprint space. To maintain spatial consistency, we introduce a displacement compensation $\Delta u$ along the $u$-axis to account for instantaneous contact point changes during finger rolling. Let $l$ be the original cross-sectional arc and $l_m$ be the arc after rotation by angle $\theta$, both unfolded to the $u$-axis using the same parameterization. The compensation is:
\begin{equation}
\Delta u = \min(u|{l_m}^{\prime}) - \min(u|{l^{\prime}})
\end{equation}
The final projection position is:
\begin{equation}
{l_m^{p}}^{\prime} = l_m^p + \Delta u
\end{equation}
Fig.~\ref{fig:delta_u_proj} illustrates the principles of $\Delta u$ calculation and the orthographic projection correction mechanism.
\begin{figure}[!t]
\centering
\subfloat[$\Delta u$ calculation principle \label{fig:delta_u}]{\includegraphics[height=0.3\linewidth]{figures/uv_flatting_ex.pdf}} \hfil
\subfloat[Orthographic correction \label{fig:proj_u}]{\includegraphics[height=0.3\linewidth]{figures/proj_ex.pdf}}
\caption{Illustration of the $\Delta u$ displacement compensation and projection correction mechanism.}
\label{fig:delta_u_proj}
\end{figure}
This correction mechanism ensures that multi-pose contactless images, despite exhibiting nonlinear visual distortions, remain strictly aligned with the original rolled fingerprint in the coordinate system. Fig.~\ref{fig:gen_examples} presents representative samples from the full IMPOSE generation pipeline.
\begin{figure*}[!t]
\centering
\includegraphics[width=0.95\textwidth]{figures/generated_samples.png}
\caption{Representative fingerprint samples from the IMPOSE generation framework, spanning rolled fingerprints, contactless translations, and multi-pose projections.}
\label{fig:gen_examples}
\end{figure*}
\subsection{Implementation Details}
\label{sec:impl}
The generation architecture consists of hierarchical autoencoders (VQ-VAE) and latent diffusion denoising networks (LDM). The autoencoder employs a ResNet-based hierarchical design with two consecutive downsampling stages achieving $4\times$ spatial compression (from $512$ to $128$, base channel width 64). The Sauvola binary map encoder follows the VQ-VAE architecture, while the ControlNet branch maintains structural consistency with the LDM encoder. The diffusion backbone uses a symmetric U-Net structure with 8-head cross-attention at the $32 \times 32$ scale for multi-scale global context modeling.
Table~\ref{tab:training_config} summarizes the training configurations for each sub-module. During joint training of LDM and ControlNet, a 500-step learning rate warmup is employed, where the learning rate scaling factor increases linearly from $1 \times 10^{-2}$ to $1$.
\begin{table}[!t]
\caption{Training Configuration of IMPOSE Sub-Modules}
\label{tab:training_config}
\centering
\begin{tabular}{lcccc}
\toprule
\textbf{Module} & \textbf{Batch} & \textbf{LR} & \textbf{Epochs} & \textbf{GPUs} \\
\midrule
Rolled VQ-VAE & 8 & $7\times10^{-5}$ & 96 & 4$\times$RTX 4090 \\
Rolled LDM & 24 & $2\times10^{-5}$ & 780 & 4$\times$RTX 4090 \\
\midrule
Sauvola AE & 16 & $3.5\times10^{-4}$ & 120 & 2$\times$RTX 4090 \\
\midrule
CL VQ-VAE & 8 & $7\times10^{-5}$ & 60 & 4$\times$RTX 4090 \\
CL LDM & 20 & $5\times10^{-6}$ & 310 & 4$\times$RTX 4090 \\
CL ControlNet & 24 & $1\times10^{-6}$ & 360 & 4$\times$RTX 4090 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table*}[!t]
\centering
\caption{Summary of Datasets Used in IMPOSE Experiments}
\label{tab:datasets}
\begin{threeparttable}
\begin{tabular}{p{0.15\textwidth}p{0.45\textwidth}p{0.25\textwidth}}
\toprule
\textbf{Dataset} & \textbf{Description} & \textbf{Usage} \\
\midrule
NIST SD302~\cite{nist302} & 2,000 rolled fingerprints, optical sensor & Identity generation training \\
THU Rolled & 1,093 rolled fingerprints, optical sensor & Identity generation training \\
THU Finger3D & 285 3D fingerprints, structured-light scanner & Texture mapping reference \\
\midrule
UWA~\cite{zhou2014benchmark}\tnote{$\dagger$} & 1,500 fingers $\times$ 3 contactless, $\times$ 2 flattened CL, 1,500 plain & Cross-modal training\tnote{$\ddagger$} \& testing \\
PolyU CL2CB~\cite{liu20143d} & 336 fingers $\times$ 6 contact/contactless & Testing \\
\bottomrule
\end{tabular}
\begin{tablenotes}
\item[$\dagger$] Contactless samples were pre-processed with white backgrounds and CLAHE-based contrast enhancement.
\item[$\ddagger$] Only flattened images used for style training; no test identity leakage.
\end{tablenotes}
\end{threeparttable}
\end{table*}
\begin{figure*}[!t]
\centering
\subfloat[NIST SD302 rolled]{\includegraphics[width=0.31\textwidth]{figures/nist302.png}} \hfill
\subfloat[THU Rolled]{\includegraphics[width=0.31\textwidth]{figures/thu_rolled.png}} \hfill
\subfloat[THU Finger3D]{\includegraphics[width=0.31\textwidth]{figures/thu_finger3d.png}} \\
\subfloat[UWA contactless\label{fig:uwa_cl}]{\includegraphics[width=0.31\textwidth]{figures/uwa_cl.png}} \hfill
\subfloat[UWA flattened CL\label{fig:uwa_fl}]{\includegraphics[width=0.31\textwidth]{figures/uwa_flatten_cl.png}} \hfill
\subfloat[UWA plain\label{fig:uwa_plain}]{\includegraphics[width=0.31\textwidth]{figures/uwa_cb.png}}
\caption{Representative fingerprint samples from the datasets used in IMPOSE experiments.}
\label{fig:dataset_samples}
\end{figure*}
During autoencoder pretraining, aggressive geometric augmentation is applied (random translation within $\pm 52$ pixels, rotation within $\pm 45^\circ$, horizontal flipping) to strengthen representation robustness. For subsequent LDM and ControlNet training, only horizontal flipping is retained while translation and rotation are strictly disabled to ensure generated images maintain spatial consistency with the standard fingerprint coordinate system.
For quality assurance, generated rolled fingerprints are filtered using NFIQ 2.0~\cite{tabassi2021nist} (quality score $> 0.55$) and foreground ratio ($> 60\%$). The dataset is split 7:3 for training and validation; after optimal hyperparameter selection on the validation set, the full dataset is used for final model training.
\section{Experimental Datasets}
The experimental datasets encompass three key categories: high-quality rolled fingerprints for identity generation training, contactless fingerprints for style transfer training and evaluation, and 3D fingerprints for multi-pose simulation. All 2D fingerprint images are normalized to 500 PPI. Table~\ref{tab:datasets} summarizes the dataset details.
For rolled fingerprint generation, we train on 3,093 images from NIST SD302~\cite{nist302} and THU Rolled. Compared to existing methods~\cite{engelsma2022printsgan,grosz2024universal}, IMPOSE achieves rich identity output with lower data dependency.
For cross-modal training, UWA flattened contactless images are used for style transfer. Critically, the training only learns modality style distributions---the generated fingerprint identities originate from the independent rolled fingerprint generator, thus eliminating test identity leakage risk. Multi-pose simulation uses the THU Finger3D dataset by randomly selecting 3D point cloud models as geometric templates.
For evaluation, we adopt the protocol established by Cui et al.~\cite{cui2023monocular}. For the UWA dataset~\cite{zhou2014benchmark}, the gallery is formed by the initial contact-based plain fingerprints of the first 1,000 fingers, while the query set consists of the corresponding contactless images captured at three angles. For the PolyU CL2CB dataset ~\cite{liu20143d}, contactless fingerprints are pre-processed by scaling them to the mean ridge period of the contact-based samples, notably omitting any geometric warping.
\section{Experimental Results}
\subsection{Rolled Fingerprint Generation Evaluation}
To quantitatively verify the rolled fingerprint generation module's ability to capture the target domain distribution, we randomly sample 3,000 synthetic rolled fingerprints (comparable to the 3,093 training samples).
\textbf{Image Quality Distribution.} We employ NFIQ 2.0~\cite{tabassi2021nist} to analyze quality distribution alignment between synthetic and real samples. As shown in Fig.~\ref{fig:nfiq}, the synthetic distribution closely mirrors the training set, with mean NFIQ 2.0 scores of 47.59 (synthetic) vs.\ 52.81 (real) and standard deviations of 24.51 vs.\ 20.94. The quality distribution peaks and main coverage intervals are highly consistent, validating IMPOSE's accurate capture of real rolled fingerprint quality statistics.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{figures/nfiq_dist_en.png}
\caption{NFIQ 2.0 quality distribution comparison between synthetic rolled fingerprints and real training samples.}
\label{fig:nfiq}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{figures/four_score_dist_en.png}
\caption{Match score distributions evaluating identity consistency, uniqueness, and privacy preservation of synthetic rolled fingerprints.}
\label{fig:match_score}
\end{figure}
\begin{figure*}[!t]
\centering
\includegraphics[width=0.95\textwidth]{figures/id_consistent_shown.pdf}
\caption{Qualitative comparison of identity consistency under different data alignment strategies for cross-modal fingerprint generation. The baseline model (rigid alignment) exhibits ridge structure collapse, while IMPOSE (Sauvola-guided) faithfully preserves minutiae topology.}
\label{fig:id_control}
\end{figure*}
\textbf{Identity Properties.} Using FDD~\cite{pan2024FDD} as the matching tool, we evaluate three dimensions: (1) genuine/impostor baseline on real data, (2) privacy leakage (synthetic vs.\ training samples), and (3) identity uniqueness (synthetic vs.\ synthetic). As shown in Fig.~\ref{fig:match_score}, synthetic-training match scores fall entirely within the impostor distribution, confirming no training identity leakage. Synthetic-synthetic match scores also follow an impostor distribution, demonstrating that the module generates distinct, unique fingerprint identities.
\subsection{Contactless Fingerprint Generation Evaluation}
\label{sec:cl_gen_eval}
We qualitatively assess identity topology preservation during cross-modal generation. As emphasized in Section~\ref{sec:sauvola}, IMPOSE employs ridge-level elastically aligned Sauvola binary maps as structural anchors, enabling pixel-precision ridge topology transfer without additional identity supervision.
\textbf{Identity Consistency Analysis.} A baseline model using conventional rigid alignment (VeriFinger~\cite{nist2020verifinger} binarization after minutiae-based global rigid transformation) is trained for comparison. Fig.~\ref{fig:id_control} visualizes the results. The baseline model, constrained by rigid alignment of training pairs, cannot learn complex contactless optical deformation patterns, resulting in severe ridge structure collapse or hallucinated textures in high-deformation regions. In contrast, IMPOSE precisely reproduces the ridge trajectory and minutiae topology of the source binary map, with dense and accurate matching correspondences.
\textbf{Cross-Database Modality Generalization.} IMPOSE's training logic exhibits strong modality transferability. Fig.~\ref{fig:polyu_gen} shows modality transfer results on PolyU CL2CB. The generated samples accurately reproduce PolyU CL2CB's characteristic localized distribution and low-contrast visual style while maintaining ridge-level identity consistency, demonstrating robust generation capability across diverse contactless modalities.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{figures/gen_style_compare.png}
\caption{Generation results of IMPOSE under different contactless fingerprint styles (UWA and PolyU CL2CB).}
\label{fig:polyu_gen}
\end{figure}
\subsection{FDD Performance Enhancement with IMPOSE Data}
\label{sec:fdd_enhance}
We evaluate the impact of IMPOSE-synthesized data on FDD~\cite{pan2024FDD}, which consists of a fingerprint pose estimation module~\cite{guan2025finger} and a dense descriptor extraction module. Rank-1 identification rate, TAR@FAR=0.1\%, and Equal Error Rate (EER) serve as evaluation metrics.
\textbf{Experimental Setup.} From the rolled fingerprint generator, 10,008 identities are randomly sampled, of which 1,712 high-quality samples pass quality filtering. Through modality translation and multi-pose simulation, each identity generates 9 images: 1 frontal, 4 positive roll angles ($0^\circ$--$60^\circ$), and 4 negative roll angles ($-60^\circ$--$0^\circ$), yielding $1,712 \times 9 = 15,408$ synthetic contactless fingerprints.
The baseline uses models trained solely on real data. Fine-tuning is performed on the pose estimation module (3 epochs) and descriptor module (5 epochs) separately and jointly.
\begin{table*}[!t]
\centering
\caption{Cross-Modal Matching Performance of FDD on UWA Database (\%)}
\label{tab:fdd_uwa}
\resizebox{0.95\textwidth}{!}{
\begin{threeparttable}
\begin{tabular}{lccccccccc}
\toprule
\multirow{2}{*}{\textbf{Method}} & \multicolumn{3}{c}{\textbf{Left}} & \multicolumn{3}{c}{\textbf{Front}} & \multicolumn{3}{c}{\textbf{Right}} \\
\cmidrule(lr){2-4} \cmidrule(lr){5-7} \cmidrule(lr){8-10}
& \textbf{R-1} & \textbf{TAR} & \textbf{EER} & \textbf{R-1} & \textbf{TAR} & \textbf{EER} & \textbf{R-1} & \textbf{TAR} & \textbf{EER} \\
\midrule
Cui et al.~\cite{cui2023monocular} & -- & -- & 16.01 & -- & -- & 3.59 & -- & -- & 16.47 \\
\midrule
FDD~\cite{pan2024FDD} baseline & 60.34 & 59.94 & 15.06 & 95.78 & 95.58 & 2.19 & 63.86 & 63.25 & 17.17 \\
FT Descriptor & 59.54 & 58.23 & 16.87 & 96.99 & 96.89 & 1.71 & 67.07 & 65.26 & 16.57 \\
FT Pose & 68.88 & 68.88 & 11.67 & 95.68 & 94.78 & 2.81 & 71.89 & 72.79 & 12.85 \\
\rowcolor{gray!15}
FT Pose $+$ Descriptor & 69.96 & 69.88 & \textbf{10.94} & 96.79 & 96.18 & \textbf{2.21} & 73.49 & 73.49 & \textbf{12.63} \\
\bottomrule
\end{tabular}
\begin{tablenotes}
\item[$*$] TAR = TAR@FAR=0.1\%. R-1 = Rank-1. FT = Fine-Tuned.
\end{tablenotes}
\end{threeparttable}}
\end{table*}
\textbf{Quantitative Results.} Table~\ref{tab:fdd_uwa} reports results on UWA partitioned into left, front, and right roll subsets. The baseline FDD degrades significantly under large roll angles, with the pose estimation module exhibiting substantial geometric bias on non-central regions with severe nonlinear distortion. Fine-tuning the pose module with IMPOSE data dramatically improves prediction accuracy at extreme angles, reducing EER from 15.06\% to 11.67\% (left) and from 17.17\% to 12.85\% (right). Fine-tuning the descriptor module benefits the front subset (EER from 2.19\% to 1.71\%), reflecting enhanced modality-specific feature extraction. Joint fine-tuning achieves the optimal overall results. As shown in Fig.~\ref{fig:pose_aug}, fine-tuning the pose estimation module with IMPOSE data significantly improves geometric prediction accuracy for large-angle contactless fingerprints, while preserving robust estimation on standard contact-based fingerprints.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{figures/pose_augment_examples.png}
\caption{Visualization of pose estimation results before and after fine-tuning with IMPOSE data on large-angle contactless fingerprints.}
\label{fig:pose_aug}
\end{figure}
Fig.~\ref{fig:uwa_cmc_det} presents comprehensive CMC and DET curves for all fine-tuning strategies, confirming that jointly fine-tuning both modules achieves superior performance across all evaluation subsets.
\begin{figure}[!t]
\centering
\subfloat[CMC curves]{\includegraphics[height=0.4\linewidth]{figures/cmc_UWA_all.png}} \hfil
\subfloat[DET curves]{\includegraphics[height=0.4\linewidth]{figures/det_UWA_all.png}}
\caption{CMC and DET curves of FDD under different fine-tuning strategies on the UWA database (all subsets combined).}
\label{fig:uwa_cmc_det}
\end{figure}
Compared to the monocular depth-based unwarping method of Cui et al.~\cite{cui2023monocular}, FDD with IMPOSE data achieves superior performance without explicit geometric preprocessing, demonstrating the efficacy of learning implicit nonlinear mappings through physically-aligned synthetic data.
\subsection{Comprehensive Comparison with State-of-the-Art}
Table~\ref{tab:comparison} compares the jointly fine-tuned FDD against recent state-of-the-art methods including Cui et al.~\cite{cui2023monocular}, Jia et al.~\cite{jia2024improving}, and Dong et al.~\cite{dong2025bridging}, all of which follow the ``explicit geometric correction'' paradigm. Performance metrics are cited from their original papers for protocols consistent with ours.
\begin{table}[!t]
\centering
\caption{Cross-Modal Matching Comparison on UWA and PolyU CL2CB (\%)}
\label{tab:comparison}
\begin{threeparttable}
\begin{tabular}{lcccccc}
\toprule
\multirow{2}{*}{\textbf{Method}} & \multicolumn{3}{c}{\textbf{UWA}} & \multicolumn{3}{c}{\textbf{PolyU CL2CB}} \\
\cmidrule{2-4} \cmidrule{5-7}
& \textbf{R-1} & \textbf{TAR} & \textbf{EER} & \textbf{R-1} & \textbf{TAR} & \textbf{EER} \\
\midrule
Cui et al.~\cite{cui2023monocular} & -- & -- & 12.72 & -- & -- & 5.00 \\
Jia et al.~\cite{jia2024improving} & -- & -- & 11.16 & -- & -- & -- \\
Dong et al.~\cite{dong2025bridging} & -- & -- & 9.93 & -- & -- & 5.86 \\
\midrule
FDD~\cite{pan2024FDD} & 73.33 & 73.39 & 12.01 & 91.82 & 88.17 & 3.20 \\
Fine-tuned FDD & 80.09 & 80.29 & \textbf{8.74} & 95.68 & 92.37 & \textbf{2.26} \\
\bottomrule
\end{tabular}
\begin{tablenotes}
\item[$*$] TAR = TAR@FAR=0.1\%.
\end{tablenotes}
\end{threeparttable}
\end{table}
On UWA (all poses combined), the fine-tuned FDD achieves 8.74\% EER, substantially outperforming Dong et al.~\cite{dong2025bridging} (9.93\%), Jia et al.~\cite{jia2024improving} (11.16\%), and Cui et al.~\cite{cui2023monocular} (12.72\%). On PolyU CL2CB, the fine-tuned FDD reaches 95.68\% Rank-1 and 2.26\% EER, far surpassing competing methods. These results validate that large-scale, physically-aligned synthetic data enables end-to-end models to surpass traditional explicit geometric correction approaches in handling cross-modal nonlinear distortions.
\subsection{Generalizability to Mainstream Fixed-Length Representations}
\label{sec:generalize}
To verify the universal value of IMPOSE data, we evaluate DeepPrint~\cite{engelsma2021deepprint} and AFRNet~\cite{grosz2023afrnet}, which rely on Spatial Transformer Networks for implicit spatial normalization rather than explicit pose correction.
\begin{table*}[!t]
\centering
\caption{Cross-Modal Matching of Mainstream Fixed-Length Representations with IMPOSE Data (\%)}
\label{tab:other_multi_pose}
\resizebox{0.95\textwidth}{!}{
\begin{threeparttable}
\begin{tabular}{lccccccccc}
\toprule
\multirow{2}{*}{\textbf{Method}} & \multicolumn{3}{c}{\textbf{Left}} & \multicolumn{3}{c}{\textbf{Front}} & \multicolumn{3}{c}{\textbf{Right}} \\
\cmidrule(lr){2-4} \cmidrule(lr){5-7} \cmidrule(lr){8-10}
& \textbf{R-1} & \textbf{TAR} & \textbf{EER} & \textbf{R-1} & \textbf{TAR} & \textbf{EER} & \textbf{R-1} & \textbf{TAR} & \textbf{EER} \\
\midrule
DeepPrint~\cite{engelsma2021deepprint} & 29.62 & 27.31 & 18.05 & 75.60 & 67.67 & 4.82 & 34.94 & 32.93 & 13.86 \\
FT DeepPrint & 51.20 & 48.69 & 13.55 & 86.35 & 85.94 & 4.52 & 55.72 & 54.72 & 12.85 \\
\midrule
AFRNet~\cite{grosz2023afrnet} & 41.47 & 41.06 & 20.68 & 91.77 & 90.46 & 4.11 & 51.91 & 51.81 & 15.86 \\
FT AFRNet & 55.91 & 53.92 & 14.66 & 91.27 & 90.66 & 3.82 & 54.22 & 54.52 & 14.67 \\
\bottomrule
\end{tabular}
\begin{tablenotes}
\item[$*$] TAR = TAR@FAR=0.1\%. FT = Fine-Tuned.
\end{tablenotes}
\end{threeparttable}}
\end{table*}
Table~\ref{tab:other_multi_pose} shows consistent performance gains across all evaluation subsets after fine-tuning with IMPOSE data. Gains are most pronounced at large roll angles: DeepPrint's Rank-1 on the left subset jumps from 29.62\% to 51.20\%, and on the right from 34.94\% to 55.72\%. Even on the front subset where baseline performance is already competitive, fine-tuning yields consistent EER improvements (AFRNet: 4.11\% to 3.82\%). These results demonstrate that IMPOSE data universally enhances CNN-based representations for contactless fingerprint texture patterns, regardless of architectural differences.
\subsection{Synthetic vs.\ Real Data Fine-Tuning}
\label{sec:syn_vs_real}
To investigate the substitution potential of IMPOSE data, we compare fine-tuning with synthetic data versus real contactless data. Real training data uses 500 UWA fingers (outside the test set), comprising $500 \times 4 = 2,000$ plain fingerprints and $500 \times 3 \times 2 = 3,000$ contactless images. DeepPrint and AFRNet, which have fewer spatial alignment constraints, serve as evaluation backbones.
\begin{table}[!t]
\centering
\caption{Performance Comparison under Different Fine-Tuning Strategies (\%)}
\label{tab:ft_compare}
\resizebox{\linewidth}{!}{
\begin{threeparttable}
\begin{tabular}{llcccccc}
\toprule
\multirow{2}{*}{\textbf{Method}} & \multirow{2}{*}{\textbf{Strategy}} & \multicolumn{3}{c}{\textbf{UWA}} & \multicolumn{3}{c}{\textbf{PolyU CL2CB}} \\
\cmidrule{3-5} \cmidrule{6-8}
& & \textbf{R-1} & \textbf{TAR} & \textbf{EER} & \textbf{R-1} & \textbf{TAR} & \textbf{EER} \\
\midrule
\multirow{4}{*}{DeepPrint~\cite{engelsma2021deepprint}} & Baseline & 46.72 & 42.14 & 12.99 & 80.65 & 53.41 & 5.07 \\
& IMPOSE FT & 64.42 & 62.85 & 10.85 & 93.30 & 81.43 & 3.79 \\
& UWA FT & 64.93 & 62.99 & 9.10 & 92.61 & 80.23 & 3.22 \\
& Hybrid FT & 68.37 & 65.96 & \textbf{7.80} & 94.84 & 84.87 & \textbf{2.22} \\
\midrule
\multirow{4}{*}{AFRNet~\cite{grosz2023afrnet}} & Baseline & 61.71 & 60.98 & 14.46 & 77.13 & 70.14 & 10.99 \\
& IMPOSE FT & 64.42 & 62.85 & 10.85 & 93.30 & 81.43 & 3.79 \\
& UWA FT & 64.29 & 63.39 & 10.78 & 78.82 & 70.47 & 6.39 \\
& Hybrid FT & 70.18 & 70.88 & \textbf{9.44} & 87.20 & 80.14 & \textbf{4.80} \\
\bottomrule
\end{tabular}
\begin{tablenotes}
\item[$*$] TAR = TAR@FAR=0.1\%. FT = Fine-Tuned.
\end{tablenotes}
\end{threeparttable}}
\end{table}
Table~\ref{tab:ft_compare} compares four training strategies: baseline, IMPOSE-only fine-tuning, UWA real-data fine-tuning, and hybrid fine-tuning. IMPOSE-only fine-tuning achieves dramatic improvements over baselines across both databases. Notably, on cross-database PolyU CL2CB evaluation, IMPOSE fine-tuning even outperforms real UWA data fine-tuning for AFRNet, demonstrating superior out-of-domain generalization---a key advantage of physically-aligned synthetic data that provides more universal feature constraints, mitigating overfitting to the real data's specific acquisition environment.
Hybrid fine-tuning (IMPOSE + UWA) consistently achieves the best results. For DeepPrint on UWA, the hybrid strategy reduces EER to 7.80\%, significantly outperforming single-source fine-tuning. Fig.~\ref{fig:ft_det_cmc_dp} and ~\ref{fig:ft_det_cmc_afrnet} present the CMC curves across different fine-tuning strategies, confirming the consistent superiority of the hybrid approach. The combination of IMPOSE's precise physical annotations and real samples' natural imaging characteristics enables deep feature fusion without distribution shift, comprehensively capturing discriminative contactless fingerprint features.
\begin{figure}[!t]
\centering
\subfloat[UWA]{\includegraphics[height=0.4\linewidth]{figures/cmc_UNSW.png}} \hfil
\subfloat[PolyU CL2CB]{\includegraphics[height=0.4\linewidth]{figures/cmc_PolyU.png}}
\caption{CMC curves of DeepPrint under different fine-tuning strategies.}
\label{fig:ft_det_cmc_dp}
\end{figure}
\begin{figure}[!t]
\centering
\subfloat[UWA]{\includegraphics[height=0.4\linewidth]{figures/cmc_UNSW_AFRNet.png}} \hfil
\subfloat[PolyU CL2CB]{\includegraphics[height=0.4\linewidth]{figures/cmc_PolyU_AFRNet.png}}
\caption{CMC curves of AFRNet under different fine-tuning strategies.}
\label{fig:ft_det_cmc_afrnet}
\end{figure}
\section{Discussion and Limitations}
The experimental results validate IMPOSE's efficacy in significantly enhancing fixed-length dense representations (FDD~\cite{pan2024FDD}) and mainstream descriptors including DeepPrint~\cite{engelsma2021deepprint} and AFRNet~\cite{grosz2023afrnet}. However, several limitations warrant discussion.
\textbf{Pose Coverage.} Our study focuses on the roll angle, which is the primary challenge in contactless fingerprint recognition. In practice, finger pose exhibits multi-degree-of-freedom characteristics---pitch variations also introduce nonlinear projection distortions and texture compression that remain unexplored, primarily due to the lack of mature, publicly available multi-pose evaluation benchmarks covering full degree-of-freedom annotations. The IMPOSE simulation framework is inherently extensible to full-dimensional pose synthesis by incorporating pitch rotation of 3D finger models, which constitutes an important future direction.
\textbf{3D Model Diversity.} The simulation relies on existing 3D fingerprint databases with relatively fixed surface statistics, lacking detailed modeling of finger anatomical structures and skin deformation properties. Future work may explore generative techniques to synthesize more diverse 3D fingerprint meshes, and investigate end-to-end diffusion-based synthesis frameworks that combine coarse geometric features from physics engines with deep generative models, producing contactless images with more realistic texture details and natural illumination.
\textbf{From Simulation to Generation.} A promising evolution direction is transitioning from explicit geometric rendering toward end-to-end generation. By integrating physics-captured geometric features with diffusion models, the pipeline could generate contactless images with richer texture and natural illumination, further unlocking synthetic data's potential for representation enhancement in complex scenarios.
\section{Conclusion}
This paper presented IMPOSE, a physics-inspired framework for identity-consistent multi-pose generation of contactless fingerprints. IMPOSE establishes a closed-loop pipeline spanning rolled fingerprint identity mapping, cross-modal contactless image generation, and 3D-geometry-based multi-pose simulation, fundamentally ensuring rigorous ridge-level identity consistency and natural texture representation. The generated 2D samples are strictly confined to the standard fingerprint pose space with pixel-level physical alignment annotations, making them highly compatible with the training requirements of fixed-length dense representations.
Comprehensive experiments validated that synthetic data effectively aligns with real physical scenarios in feature distribution, identity independence, and ridge minutiae consistency. Fine-tuning across diverse representation architectures demonstrated IMPOSE's significant performance enhancement for FDD, DeepPrint, and AFRNet, confirming universal cross-architecture benefits. Comparisons with real data fine-tuning further established the unique value of high-quality synthetic data in feature constraints and out-of-domain generalization, demonstrating strong substitution potential in scenarios with scarce real samples.
\bibliographystyle{IEEEtran}
\bibliography{CL_generation,descriptor,others,preprocessing}
\end{document}