An over-complete sparse representation approach for face recognition under partial occlusion

B-Joint Sparsity Model (B-JSM) was presented for expression-invariant face recognition by Pradeep Nagesh and Baoxin Li in 2009, which can save storage space for grossly representing per class training images of a given subject by only two features and performs better than the state-of-the-art algorithm. But the recognition rate (RR) is very low by B-JSM when certain part is occluded. On the basis of B-JSM, a new improved model is presented to recognize human faces under partial occlusion in this paper. Firstly, we introduce B-JSM theory. Then we analyze the reason and B-JSM is improved: the feature is extracted after “getting rid of” the region containing the maximal information. A series of experiments with the Extended Yale B database show that our improved approach is effective to solve the problem of partial occlusion and robust to the low-dimensional image or only a few images of an individual.