CNN-Based Dorsal Hand Vein Authentication Using Triplet Loss Metric Learning

Abstract: Hand vein biometrics has emerged as a reliable authentication modality due to the uniqueness and internal nature of vascular patterns. Unlike surface-based biometrics such as fingerprints or facial recognition, vein patterns are less susceptible to environmental variations and spoofing attacks. This work presents a CNN-based dorsal hand vein biometric authentication framework that combines a ten-stage preprocessing pipeline with deep metric learning-based feature extraction. The preprocessing pipeline includes Gaussian smoothing, automatic hand region detection, wrist removal using distance-transform thickness profiling, finger removal using convex hull-based techniques, Otsu segmentation, CLAHE contrast enhancement, Sato vesselness filtering, feathered mask application, and tight crop resize to 224×224 pixels to isolate the stable metacarpal vein region.

A four-block CNN backbone is trained using online semi-hard triplet loss mining with an identity-based batch sampler to generate discriminative 128-dimensional L2-normalized embeddings for biometric matching. A manually collected dataset of 5220 dorsal hand vein images from 261 individuals (522 hand identities, with left and right hands treated as separate biometric identities) captured under controlled near-infrared illumination was used for training and evaluation. Experimental results demonstrate a recognition accuracy of 99.12% and an Equal Error Rate of 1.32%, confirming the effectiveness of the proposed framework for secure hand vein biometric authentication.

Keywords: Authentication, Biometric Recognition, Convolutional Neural Network (CNN), Dorsal Hand Vein Recognition, Embedding, Metric Learning, Near-Infrared Imaging, Online Semi-Hard Mining, Triplet Loss, Vessel Enhancement.