Performance Comparison of PCANet-Based Deep Learning Techniques for Palmprint Recognition

A bstract-Palmprint recognition is one the hand-based biometric technology that utilizes the unique patterns of palmprint image for recognition or identification tasks. The recognition accuracy of biometric systems is mostly relying on the distinctive features extracted from features extraction stage using algorithms. Deep learning technique is efficient technique for feature extraction and has received much attention because its ability to learn from data. Basically, deep learning uses convolutional architecture to discover multiple level of representations in which higher level of features represent more useful information. In this paper, performance comparison of several simple deep learning techniques on palmprint recognition is presented. Experiments were conducted on region of interest (ROI) of palmprint image from Tongji database using PCANet-based deep learning techniques (pooling raw, PCANet, PCANet+, PCANet with new filter) and two classic feature extraction techniques (PCA and 2DPCA). Experimental results reveal that PCAN et outperforms other deep learning and classic techniques by achieving 99.77% of recognition accuracy. A simple PCANet that consists of only a cascaded linear map followed by a nonlinear output stage offers effective and outstanding performance. It demonstrates that the PCANet is one of the deep learning techniques that capable of extracting significant information for the palmprint recognition.