Automatic region segmentation for high-resolution palmprint recognition: Towards forensic scenarios

Abstract

Recently, a novel matching strategy based on regional fusion for high resolution palmprint recognition arises for both forensic and civil applications, under the concept of different regional discriminability of three palm regions, i.e., interdigital, hypothenar and thenar. This matching strategy requires accurate automatic region segmentation techniques since manual region segmentation is time consuming. In this work, we develop automatic region segmentation techniques based on datum point detection for high-resolution palmprint recognition which can be further applied to forensic applications. Firstly, Canny edge detector is applied to a full palmprint to obtain gradient magnitudes and strong edges. Then a first datum point, i.e., the endpoint of heart line, is detected by using convex hull on gradient magnitude image and its left/right differential image and strong edge image. A second datum point, i.e., the endpoint of life line, is estimated based on the position and direction of the first datum point and statistical average distance between the two datum points. Finally, segmented palm regions are generated based on the two datum points and their perpendicular bisector. To evaluate the accuracy of our region segmentation method, we compare the automatic segmentation with manual segmentation performed on a public high resolution palmprint database THUPALMLAB with full palmprint images. The regional error rates of interdigital, thenar and hypothenar regions are 15.72%, 17.05% and 21.38% respectively. And the total error rate is 19.54% relative to full palmprint images.