Ear recognition under random occlusion via non-negative sparse representation

One challenging problem inevitable in real application is that the ears are often occluded by some objects such as hair or hat. In this paper, a general classification algorithm based on non-negative sparse representation is proposed to handle ear recognition under random occlusion. Unlike sparse representation based classification in which the input data are described as a combination of basis features involving both additive and subtractive components, the proposed classification paradigm expresses an input ear signal as a linear additive combination of all the training ear signals, and then classification is made according to the reconstruction error of the input ear image. The recognition performance for various levels of occlusion areas is investigated in which the location of occlusion is randomly chosen to simulate real scenario. Experimental results on USTB ear database reveal that when the ear is occluded, the proposed method exhibits great robustness and achieves better recognition performance.