Towards Cross-Dataset Palmprint Recognition Via Joint Pixel and Feature Alignment

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Deep learning palmprint recognition algorithms typically perform well within single datasets.
• Existing methods struggle with cross-dataset scenarios, limiting practical applications with diverse data sources like smartphones and embedded terminals.

Purpose of the Study:

• To propose a novel Joint Pixel and Feature Alignment (JPFA) framework for effective cross-dataset palmprint recognition.
• To address the challenge of domain shift in palmprint data collected from different sources.

Main Methods:

• A two-stage alignment process involving deep style transfer for pixel-level adaptation and a deep domain adaptation model for feature-level alignment.
• Utilizing deep style transfer to generate synthetic data, reducing dataset discrepancies and augmenting the training set.
• Developing a new deep domain adaptation model to align feature distributions between source and target datasets.

Main Results:

• The proposed JPFA framework significantly outperforms existing models on cross-dataset palmprint recognition tasks.
• Achieved up to a 28.10% improvement in cross-dataset identification accuracy compared to baseline methods.
• Reduced the Equal Error Rate (EER) for cross-dataset verification by up to 4.69%.

Conclusions:

• The JPFA framework offers a robust solution for cross-dataset palmprint recognition, overcoming limitations of single-dataset training.
• Demonstrated state-of-the-art performance, highlighting the effectiveness of joint pixel and feature alignment strategies.
• The study provides reproducible results with publicly available code.