Multiscale representation for partial face recognition under near infrared illumination

Abstract

Near infrared (NIR) partial face images acquired in iris recognition in less constrained environments contain plentiful identity information which have not been fully exploited. In this paper, a NIR partial face recognition (PFR) algorithm is designed according to the characteristics of NIR partial face images acquired in iris recognition systems. In the preprocessing stage, the eye corners are utilized for alignment because the eye regions are usually visible. In the feature representation stage, highly compact and discriminatory features are extracted based on Multiscale Double Supervision Convolutional Neural Network (MD-SCNN) from several multiscale patches which are cropped from the aligned images. Weights of each multiscale patch are learned to improve the performance of PFR. Finally, the dissimilarity between two partial face images is calculated as the weighted t <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> distance between corresponding patches. A new NIR partial face (NIR-PF) database is constructed for extensive evaluation, which includes 5300 images acquired from 276 subjects when they walk towards an iris imaging device. Experimental results on the NIR-PF and CASIA-IrisV4-Distance databases demonstrate the effectiveness and efficiency of the proposed algorithm.