PVDiff: A Palm-Vein Data Augmentation Method Based on Diffusion Model

Palm-vein recognition has garnered significant attention recently due to its superior privacy and safety features. However, the public dataset for palm-vein is much less. Moreover, manually collecting palm-vein data from diverse individuals is challenging and complex. To tackle these problems, we propose a palm-vein data augmentation method based on diffusion model called PVDiff. We firstly introduce the diffusion model to effectively capture the distinct spatial distributions of palm veins, so that the brand-new palm-vein images can be effectively generated. Besides, we propose several augmentation strategies for integrating the generated palm-veins to a complete dataset with good diversity and robustness. Experiment results represent that our PVDiff can not only generate high quality palm-vein data compared with other generative methods, but also can effectively enhance the performance of the recognition model. For the same backbone, the augmented data supports the model mostly decreases 78.1% on the Fail Accept Rate at Fail Recognize Rate under 0.0001%. Furthermore, our PVDiff significantly benefit the robustness of the recognition model for the complex real-world scenes.