Diff-Palm: Realistic Palmprint Generation with Polynomial Creases and Intra-Class Variation Controllable Diffusion Models

TL;DR

Diff-Palm tackles palmprint recognition data scarcity by coupling a polynomial crease representation with a creases-conditioned diffusion model. The polynomial creases capture real crease distributions via a multivariate Gaussian over coefficients, while the $K$-step noise-sharing sampling provides controllable intra-class variation. Experiments show that recognition models trained solely on Diff-Palm-generated data can surpass those trained on real data without fine-tuning, with gains increasing as the number of synthetic identities grows. The approach enables scalable, privacy-preserving palmprint synthesis and demonstrates the broader utility of crease-conditioned diffusion for biometric data generation.

Abstract

Palmprint recognition is significantly limited by the lack of large-scale publicly available datasets. Previous methods have adopted Bézier curves to simulate the palm creases, which then serve as input for conditional GANs to generate realistic palmprints. However, without employing real data fine-tuning, the performance of the recognition model trained on these synthetic datasets would drastically decline, indicating a large gap between generated and real palmprints. This is primarily due to the utilization of an inaccurate palm crease representation and challenges in balancing intra-class variation with identity consistency. To address this, we introduce a polynomial-based palm crease representation that provides a new palm crease generation mechanism more closely aligned with the real distribution. We also propose the palm creases conditioned diffusion model with a novel intra-class variation control method. By applying our proposed $K$-step noise-sharing sampling, we are able to synthesize palmprint datasets with large intra-class variation and high identity consistency. Experimental results show that, for the first time, recognition models trained solely on our synthetic datasets, without any fine-tuning, outperform those trained on real datasets. Furthermore, our approach achieves superior recognition performance as the number of generated identities increases.

Figures (12)

• Figure 1: Comparison of (a) Bézier creases zhao2022bezierpalm and (b) proposed polynomial creases.
• Figure 2: Statistical analysis of polynomial coefficients: (a) Quantile-Quantile Plot of coefficients $a_4^1$. (b) Histogram and Kernel Density Estimate (KDE) Plot of coefficients $a_4^1$. (c) Contour Plot of coefficients $a_4^1$ and $a_3^1$.
• Figure 3: The proposed intra-class variation controllable diffusion model. (a) Training process: palm crease images, extracted from palmprints using PCEM jin2024pce, are employed as conditions and concatenated with diffused palmprint images, serving as input for the UNet. (b) Sampling process: polynomial creases, as synthetic identity, are first generated and adopted to create consistent samples. The $K$-step noise-sharing sampling is applied to obtain palmprint datasets with varying degrees of intra-class variations.
• Figure 4: Generated palmprint results under different noise-sharing strategies. The figure illustrates the outcomes of applying noise-sharing in the last $K=500$ steps (top) versus the first $K=500$ steps (bottom) of a total $T=1000$ steps during the sampling process for the same identity.
• Figure 5: Comparison score plots for synthetic datasets generated by different methods and real datasets. The genuine and imposter comparison scores are calculated with features, which are extracted by a pre-trained ArcFace model from datasets.