Ear recognition with occlusion via discrimination dictionary and occlusion dictionary based sparse representation

Abstract

In ear recognition problems, sparse representation based classification (SRC) has shown good performance. The dictionary used for sparse coding plays a key role in SRC. Traditional SRC methods mostly use the holistic features of the training samples to construct the dictionary for identification. But this will bring heavy computational load because of the large dimensionality of the dictionary. Therefore, a structured dictionary construction method is proposed in this paper, which is based on the Fisher discrimination criterion. Each atom of the dictionary has correspondence to the class labels. At the same time, the Fisher discrimination criterion is applied to get the coding coefficients, which have the characteristics of small within-class scatter and big between-class scatter. Meanwhile, considering that the ear images we get in practical application always have different rates of occlusion, we propose to learn a compact occlusion dictionary with Gabor features, since Gabor features are more robust to pose variation and partial occlusion. So the whole dictionary is composed of two parts: the Fisher discrimination dictionary and the Gabor occlusion dictionary. Then we apply this dictionary in sparse representation based classification (FGSRC). The experimental results on two ear datasets have shown that our proposed FGSRC performs better than existing ear recognition under partial occlusion based on SRC; In the case of ear recognition under natural occlusion, our method also shows good performance.