Unconstrained Ear Recognition Using Deep Scattering Wavelet Network

Abstract

There has been significant progress in the field of automatic ear recognition, wherein ear images are captured in a constrained environment. But unconstrained ear recognition have acquired less attention due to the unavailability of such a database having variations in illumination, pose, size, resolution and occlusions. It is a challenging pattern recognition problem due to large intra-class variability. In this paper, we propose a novel local descriptor for unconstrained ear recognition based on scattering wavelet network (ScatNet) to extract translation and small deformation invariant local features. The experiments conducted on a recently released unconstrained ear benchmark databases, such as Annotated Web Ears (AWE) and USTB-Helloear databases, and also on our newly created IIT-Bombay smartphone-captured ear database show the effectiveness and robustness of the proposed local feature descriptor in terms of Equal Error Rate (EER) and Rank-1 (R1) accuracy.