A Mobile Palmprint Authentication System Using a Modified MNT Algorithm, Circular Local Binary Pattern, and CNN (mobileNet)

Abstract

A few approaches have been proposed for hand segmentation in palmprint recognition. Skin-color information does not process sufficient information for discrimination in complex backgrounds and variable illumination. The use of guides has also been proposed, which restricts hand placement during capturing. Contour tracing algorithms have also been proposed in the literature. This worked in an even background scenario with no objects or patterns around the hand. In the case of uneven background with objects present, the traditional contour tracing algorithm cannot accurately segment the hand from the background. Hence, this paper proposes a modified Moore Neighbor Tracing (MNT) algorithm for hand detection and key-point extraction in complex backgrounds. The hand image is converted to grey, and the edges in the hand image are detected. The modified algorithm then transverses selected edges and returns the peak and valleys of each finger. This is then used to crop the palm. The modified algorithm improves the accuracy of hand detection in complex backgrounds with an F-Score of 0.8657. A mobile palmprint biometric system was also presented using Circular Local Binary Pattern (CLBP) and Convolutional Neural Network (CNN). The system showed an accuracy of 98.3% for hands captured with the mobile device and the CASIA online database. An accuracy of 99.0% was also recorded for GPDS and PolyU online databases.