Robust Palmprint Recognition via Multi-Stage Noisy Label Selection and Correction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Deep learning significantly advances palmprint recognition.
• Current methods struggle with noisy labels, impacting reliability in real-world applications.
• Robustness against label noise is crucial for practical palmprint recognition systems.

Purpose of the Study:

• To propose a novel framework for robust palmprint recognition under noisy label conditions.
• To enhance the reliability and performance of deep learning models for palmprint identification.
• To address the challenge of label noise in large-scale palmprint datasets.

Main Methods:

• Developed a Multi-stage Noisy Label Selection and Correction (MNLSC) framework.
• Employed self-supervised learning for initial clean sample selection.
• Utilized a Fourier-based module for clean hard sample identification.
• Implemented a prototype-based module for noisy label detection and correction.

Main Results:

• The MNLSC framework demonstrated superior performance in handling varying noise rates.
• Experimental results on multiple palmprint databases confirmed the method's effectiveness.
• Achieved up to a 33.45% accuracy improvement compared to baseline methods with 60% label noise.

Conclusions:

• The proposed MNLSC framework significantly improves the robustness of deep learning-based palmprint recognition.
• The method effectively handles noisy labels, leading to more reliable biometric identification.
• This approach offers a promising solution for real-world palmprint recognition systems facing data imperfections.