Sparse recognition via intra-class dictionary learning using visual saliency information

In recent years, sparse recognition (SR) has increasingly become an emerging pattern recognition method. Because of its excellent recognition performance for some traditionally difficult problems (such as occluded or corrupted face recognition), several classical SR ideas (such as sparse representation-based classification (SRC) or dictionary-based sparse recognition (DSR)) have been the focus of research in the intelligent information field. However, for image recognition against actual backgrounds, there are still problems with these mainstream SR methods. Hence, this paper presents a new SR method which combines the advantages of both SRC and DSR. In the pre-processing, visual saliency information (VSI) for images with complex scenes is extracted by introducing the saliency map as a tool. Then, DSR is used to develop intra-class dictionary learning for the VSI data. The last step is to solve a l1-norm optimization problem to give the SR result by generating a global recognition matrix with the SRC mechanism. Experimental results show that the proposed method for ‘real world’ image recognition provides advantages over mainstream SR methods, in recognition rate and computation time cost.