Finger knuckle recognition with PCANet

Abstract

This paper presents finger knuckle recognition system using feature extraction method, Principal Component Analysis Network (PCANet). Finger knuckle is one of the most secure traits in biometric system and convenient for personal recognition. In this work, a very simple deep learning network, PCANet is employed to extract discriminative features from the original grayscale of finger knuckle image. PCANet utilises whole input knuckle image to generate the filter and images are convolved with the filter bank. The final feature vectors generated from the output of the last convolutional layer are used for the classification. Experiments were conducted on finger knuckles images from two databases (USM and THU-FDT2) to verify the performance of finger knuckle recognition system with PCANet as feature extraction method. Experimental results on two databases exhibit the robustness and superiority of the PCANet method over the other feature extraction methods (PCA, 2DPCA, Pooling raw). The findings also show that the number of filters in PCANet affects the classification accuracy of the finger knuckle recognition system. Accuracy increases as the number of filters increase. The classification accuracies using PCANet are 99.33% and 100% for finger knuckle images USM database and THU-FDT2, respectively.