Palm Vein Recognition Using Near-Infrared Imaging and CNN-Based Feature Extraction for Secure Biometric Authentication

Abstract

Biometric authentication systems have evolved significantly to meet the growing demand for secure, reliable, and scalable identity verification systems. Among the available modalities, palm vein recognition has emerged as a promising candidate due to its intrinsically secure nature-vein patterns are subcutaneous, unique to individuals, and resistant to forgery. This research investigates the design, development, and evaluation of a complete palm vein recognition system based on near-infrared imaging and convolutional neural network-driven feature learning. It integrates hardware-level acquisition, image preprocessing, feature extraction, and CNN-based matching into a unified pipeline, therefore presenting an academically rigorous investigation of the feasibility of subcutaneous biometrics assisted by deep learning.

This work is initiated with a comprehensive literature review of the state of the art in biometrics, NIR imaging physics, and computer methods for vascular pattern recognition. Traditional feature extraction methods-which include Gabor filters, Local Binary Patterns (LBP), Maximum Curvature, and Repeated Line Tracking-are reviewed to establish a relative theoretical framework. The limitations of handcrafted descriptors motivate the use of CNN-based architectures for superior hierarchical extraction of vascular features without human intervention.

A customized imaging device was designed with 850-950 nm near-infrared LEDs, an NIR-sensitive camera system, and an optical bandpass filter that can capture raw palm vein images in both controlled and semi-controlled settings. Particular attention has been paid to the optical absorption characteristics of deoxygenated hemoglobin, the scattering properties of biological tissues, and the required geometry of illumination to optimize the appearance of veins. Multiple subjects are recorded under a standardized acquisition protocol to form a structured dataset for experimentation.

Preprocessing techniques, such as histogram equalization, Gaussian smoothing, extraction of ROIs, contrast stretching, and reduction of specular noise, were used to enhance vascular clarity. A CNN classifier was then trained on this preprocessed dataset through supervised learning, thereby allowing identification by feature embeddings and classification probabilities. Even with a small database, the model demonstrated an 83% matching accuracy, highlighting the strength of deep-learning-based feature representation even under hardware constraints.

A close analysis of the system's performance, variability factors, and error distribution points to both strengths and remaining limitations of the approach. In conclusion, palm vein recognition, combined with CNN-based learning and custom NIR acquisition hardware, represents a highly promising biometric modality for applications involving high security, such as financial authentication systems, secure facility access, or integration into platforms like BioPay.

This paper provides a comprehensive, hardware-to-software contribution to the academic community related to palm vein recognition, drawing on empirical results, theoretical foundations, and practical implementation considerations. It identifies future directions, such as dataset expansion, augmentation strategies, multi-modal fusion, mobile deployment, and improved anti-spoofing mechanisms, which are meaningful pathways for continued research in subcutaneous biometrics.