Learning to Combine Local and Global Image Information for Contactless Palmprint Recognition

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Traditional palmprint recognition relies on handcrafted features, limiting performance.
• Deep learning models often extract global features, struggling with unconstrained data and elastic deformations in contactless systems.

Purpose of the Study:

• To develop a novel deep learning approach for contactless palmprint recognition that overcomes limitations of elastic deformations.
• To enhance the discriminative power of feature representations for improved palmprint identification accuracy.

Main Methods:

• A novel two-path Convolutional Neural Network (CNN) architecture was designed.
• One path processes the input holistically, while the second extracts local features from image patches.
• The model was trained using a combined Additive Angular Margin (ArcFace) Loss and center loss objective.

Main Results:

• The proposed model significantly enhances the discriminative power of learned image representations compared to standard holistic models.
• State-of-the-art performance was achieved on contactless palmprint recognition tasks.
• Favorable results were demonstrated on the IITD and CASIA contactless palmprint datasets against existing methods.

Conclusions:

• The novel two-path CNN effectively addresses elastic deformations in contactless palmprint recognition.
• The approach significantly improves accuracy and achieves state-of-the-art results.
• The publicly available source code facilitates further research and development in the field.