GenPalm: Contactless Palmprint Generation with Diffusion Models

TL;DR

The paper tackles the data scarcity and privacy constraints that hinder contactless palmprint recognition by introducing GenPalm, a two-stage diffusion-model framework for synthetic palmprint generation. Stage One fine-tunes Stable Diffusion to create diverse palm identities, while Stage Two uses an Identity-Preserving ControlNet guided by palm crease line maps and perspective distortions to produce multiple images per identity without losing identity fidelity. The approach achieves realism comparable to real palmprints and significantly enhances recognition performance when synthetic identities are combined with real data, enabling large-scale synthetic datasets (12,000 identities × 20 images) to alleviate dataset limitations. This work offers a practical path to scalable evaluation and deployment of contactless palmprint systems, with public release of a large synthetic database to support research and benchmarking.

Abstract

The scarcity of large-scale palmprint databases poses a significant bottleneck to advancements in contactless palmprint recognition. To address this, researchers have turned to synthetic data generation. While Generative Adversarial Networks (GANs) have been widely used, they suffer from instability and mode collapse. Recently, diffusion probabilistic models have emerged as a promising alternative, offering stable training and better distribution coverage. This paper introduces a novel palmprint generation method using diffusion probabilistic models, develops an end-to-end framework for synthesizing multiple palm identities, and validates the realism and utility of the generated palmprints. Experimental results demonstrate the effectiveness of our approach in generating palmprint images which enhance contactless palmprint recognition performance across several test databases utilizing challenging cross-database and time-separated evaluation protocols.

Figures (6)

• Figure 1: Example contactless palmprint region of interest (ROI) images from existing palmprint databases: (a) KTU aykut2015developing, (b) IITD v1 kumar2008incorporatingkumarpersonal, (c) MSU PalmDB grosz2024mobile, and (d) CASIA Multispectral (CASIA MS) hao2007comparativehao2008multispectral. ROIs are extracted at a resolution of 224x224 using the ROI extraction algorithm proposed in grosz2024mobile.
• Figure 2: Overview of GenPalm architecture. GenPalm is able to replicate the diversity in color and textures present in its training corpus, demonstrated in the example outputs of the ID-ControlNet module.
• Figure 3: Example GenPalm-generated images of five different synthetic identities with five images per identity. GenPalm is able to generate highly realistic palmprint ROIs with useful intra-class and inter-class separation, including various color and texture variations.
• Figure 4: Pretrained Palm-ID grosz2024mobile similarity score distributions for the real NTU-CP-v1 palmprint database and corresponding sized GenPalm-generated database. Due to the larger intra-class variations present in GenPalm images, the imposter scores and genuine score distributions are slightly more overlapped than the real dataset.
• Figure 5: T-SNE embeddings of real palmprint images and GenPalm synthetic images of the same 10 identities (denoted by different colors). The proximity of real (denoted by dots) and synthetic (denoted by stars) embeddings corresponding to the same identity (i.e., color) validates both the realism and identity preservation of GenPalm-generated images.