Generating Stylized Features for Single-Source Cross-Dataset Palmprint Recognition With Unseen Target Dataset.

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Cross-dataset palmprint recognition is crucial for real-world applications.
• Generalizing models to diverse datasets and devices remains a significant challenge.
• Existing methods struggle with unseen target datasets in single-source cross-dataset scenarios.

Purpose of the Study:

• To address the limitations of single-source cross-dataset palmprint recognition with unseen target datasets (S2CDPR-UT).
• To propose a novel method, Generating stylized Features (GIFT), to enhance feature extractor generalization.
• To improve the performance of palmprint recognition systems across different datasets and environments.

Main Methods:

• Decoupling raw features into high- and low-frequency components.
• Employing a feature stylization module to perturb low-frequency components, generating stylized features.
• Introducing diversity enhancement and consistency preservation supervisions at the feature level.

Main Results:

• The proposed GIFT method significantly improves performance in S2CDPR-UT.
• Experiments on CASIA Multi-Spectral, XJTU-UP, and MPD datasets demonstrate GIFT's effectiveness.
• GIFT expands the feature space while maintaining semantic consistency for accurate recognition.

Conclusions:

• GIFT offers a robust solution for single-source cross-dataset palmprint recognition with unseen target datasets.
• The method effectively enhances the generalization capability of palmprint feature extractors.
• The approach shows promising results for real-world biometric systems facing data variability.