Palmprint recognition based on principal line features.

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Palmprint recognition is crucial for modern security due to diverse imaging devices.
• Traditional methods struggle with effectively extracting principal line features from palmprints.
• Fine lines in palmprints can introduce noise and hinder accurate feature extraction.

Purpose of the Study:

• To introduce a novel data augmentation method for improved palmprint recognition.
• To propose a new Layer Visual Transformer (LViT) design paradigm for enhanced feature extraction.
• To evaluate the proposed methods' performance, robustness, and efficiency.

Main Methods:

• Utilized Wide Line Extraction (WLE) filter to extract principal palmprint lines based on direction and width.
• Applied Gabor filter post-WLE to purify features and remove noise from fine lines.
• Developed LViT with distinct blocking strategies for multi-level feature capture and fused results.
• Evaluated performance on multiple databases including PolyU II, IIT Delhi, XINHUA, and NTU-CP-V1.

Main Results:

• Data augmentation improved recognition rates across four Vision Transformer (ViT) models, with a 32.9% increase on the XINHUA database.
• LViT achieved state-of-the-art results on multiple databases by effectively utilizing fused local and global features.
• LViT demonstrated excellent noise-resistant generalization ability, maintaining stable performance under simulated real-world noise conditions.
• LViT required fewer training iterations compared to traditional methods.

Conclusions:

• The proposed data augmentation method significantly enhances palmprint recognition accuracy.
• LViT offers a promising new paradigm for palmprint recognition, achieving superior performance and efficiency.
• The developed approach exhibits strong robustness against noise, making it suitable for real-world applications.