An efficient deep learning framework for contactless palmprint recognition: multiresolution principal component analysis network

Contactless palmprint recognition is a widely used method of personal identification. Its performance relies primarily on the feature extraction stage, where intra-variability (pose, scale, and illumination) must be considered. This study presents a novel and challenging contactless palmprint representation called the Deep Statistical Image Features (DSIF), which combines the Discrete Wavelet Transform (DWT) with the Principal Component Analysis Network (PCANet). The methodology uses the following steps: First, the DWT of levels 1 and 2 is applied to extract different sub-band images. Next, the PCANet algorithm is applied to the palmprint image and the low-frequency sub-band images. Then, histograms are extracted and concatenated. Finally, the reduced representation is constructed using Whitened Principal Component Analysis (WPCA). The key contribution of this study is its feature extraction methodology, which uses multiresolution analysis instead of multi-patch decomposition in order to obtain pertinent information from various image resolutions. The proposed method uses the entire IIT-Delhi contactless database to construct the model, which is then tested on two other contactless palmprint databases, CASIA and Tongji. The method achieved rank-1 identification rates of 99.80 % on CASIA, 98.77 % on Right Tongji, and 99.07 % on Left Tongji, results that are impressive compared to current approaches and methods.