A Biometric System Based on Deep Learning Techniques

Abstract

Biometric authentication has become a fundamental component of modern security sys- tems, with Finger Knuckle Print (FKP) recognition emerging as a promising modality due to

its rich and stable texture patterns and low-cost acquisition. However, existing FKP systems

suffer from sensitivity to illumination variations and a lack of interpretability in deep learning- based decisions. This thesis proposes a unified, robust, and explainable framework that ad- dresses these challenges by integrating illumination-invariant preprocessing, efficient feature

extraction, and transparent classification. Specifically, the Self-Quotient Image (SQI) algo- rithm is employed to normalize illumination effects, followed by a two-stage PCANet model

to extract hierarchical and discriminative features capturing both local and global patterns. An

Explainable Deep Neural Network (xDNN) is then used for prototype-based classification, pro- viding interpretable IF–THEN rules and visual prototypes. To further enhance performance, the

framework is extended through the EDHP-FKP approach, which incorporates edge-enhanced preprocessing using Canny, Sobel, and Laplacian of Gaussian operators, evaluates multiple

deep feature extractors (VGG-16, AlexNet, ResNet-50) alongside Gabor descriptors, and in- tegrates a hierarchical prototype (HP) classifier with comparative analysis against the Nearest

Neighbor (NN) method. Extensive experiments on the PolyU FKP database demonstrate that the proposed framework achieves high recognition accuracy (up to 95.96%) with low Equal Error Rates, while ensuring robustness to illumination variations and significantly improving

interpretability compared to conventional black-box models, thereby contributing to the devel- opment of reliable and trustworthy biometric authentication systems.