Learning LogDet Divergence for Ear Recognition

Abstract

Ear-print has become one of the most important types of vital biometric in recent years; ear-print is using in different applications; especially in forensic science. In this paper, we present a novel approach for ear recognition based on fusion local descriptors for feature extraction, and LogDot divergence for classification. In details, binarized statistical image feature (BSIF) and patterns of oriented edge magnitude (POEM) are used to represent ear image. Then, discriminative correlation analysis (DCA) algorithm is exploited for fusion those features and reduction dimension. Finally, LogDot divergence based metric learning is adopted to recognize the ear images by learning a Mahalanobis matrix for approximate nearest neighbor (ANN) approach. The experimental results ar performed on four available datasets; IIT Delhi I, II and USTB I, II datasets. The proposed approach superior performance over the state-of-the-art approaches and can achieve promising recognition rates around 98.4%, 98.7%, 100% and 97.4% for IIT Delhi I, II, and USTB I, II, respectively.