Robust ear recognition using sparse representation of local features

As a local image feature description approach, LBP (local binary pattern) is regarded as one of the most effective textural features to describe images. In this paper, a general classification algorithm via sparse representation of LBP features is proposed for ear recognition. This algorithm expresses LBP features of the input ear image as a sparse combination of LBP features extracted from all the training ear images. The recognition performance for salt and pepper noise, Gaussian noise and various levels of random occlusion in which the location of occlusion is randomly chosen to simulate real scenario is investigated. Experimental results on USTB ear database reveal that when the test ear image is contaminated by noise or is occluded, the proposed approach exhibits a greater robustness and achieves a better recognition performance.