Ear Biometric Recognition in Unconstrained Conditions

Abstract

Recognizing identity from morphological shape of the human ear using one sample image per person in training-set (i.e. only one model of the individual to be identified is registered in the database and available for the task of identification), with insufficient and incomplete training data, dealing with strong person-specificity can be very challenging. In addition, most encountered testing-images in real world applications are not in high quality due to their acquisitions in difficult conditions (ex, video-surveillance) which cause more challenges like: rotated images or images with low resolution. In continuation to our previous works on ear recognition, we present in this paper an experimental and comparative study on the effects of rotation and scaling of ear images using only one sample image per person in training-set which are considered as problems largely encountered in real world applications. Several local color texture descriptors are tested and compared under several color spaces. Support Vector Machine (SVM) is used as a classifier. We experiment with USTB-1 ear database. The experiments show very acceptable and interesting results in comparison to those reported in literature.