PalmDiff: When Palmprint Generation Meets Controllable Diffusion Model.

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Palmprint recognition is crucial for biometric identity verification.
• Limited large-scale public datasets hinder palmprint research and system accuracy.
• Existing generative methods lack generalization, producing images dissimilar to conditional inputs.

Purpose of the Study:

• To propose PalmDiff, a controllable diffusion model for generating synthetic palmprint data.
• To address the challenge of insufficient palmprint datasets and improve recognition system accuracy.
• To enhance the quality and detail preservation of generated palmprint images.

Main Methods:

• Utilized a controllable diffusion model (PalmDiff) for palmprint image generation.
• Introduced a diffusion process to mitigate noise and preserve texture details.
• Employed a linear attention mechanism to boost backbone expressiveness and reduce computational load.
• Developed an ID loss function for consistent palmprint image generation within an identical space.

Main Results:

• PalmDiff demonstrated strong performance in image generation, achieving FID scores of 13.311 (MPD) and 18.434 (Tongji).
• The model effectively tackled noise and texture loss issues inherent in diffusion models.
• PalmDiff significantly improved the performance of various palmprint recognition backbones compared to other generative methods.

Conclusions:

• PalmDiff offers a viable solution for augmenting palmprint datasets and enhancing biometric recognition systems.
• The proposed method generates high-fidelity palmprint images, overcoming limitations of existing generative techniques.
• PalmDiff contributes to advancing palmprint recognition accuracy and robustness through improved data generation.