CONTACTLESS PALMPRINT RECOGNITION IN SECURITY CRITICAL ENVIRONMENTS USING END-TO-END DEEP LEARNING

Abstract

This work introduces an end-to-end deep learning model for contactless palmprint recognition that overcomes real-world variability issues like changes in illumination, background clutter, and hand position. The system combines automatic region-of-interest (ROI) extraction, pre-processing with Gabor filtering, and robust feature learning via a convolutional neural network (CNN) architecture. Feature embeddings are created and categorized based on both SoftMax and metric-based approaches such as triplet loss for top-tier identification and verification. The architecture is modular, enabling integration with other modalities such as palm vein or synthetic samples. Extreme focus on generalizability, scalability, and mathematical correctness is maintained across the pipeline. Metrics of performance such as TAR, FAR, EER, and ROC are incorporated to measure system performance. This platform provides a clean and efficient solution for biometric authentication that is appropriate for use in security-critical and public health environments where contactless engagement is required.