Visual recognition based on discriminative and collaborative representation

In this paper, a low computation complexity, yet very efficient representation of image for visual recognition tasks is presented. The collaborative representation and discriminative ingredient are combined in a unified framework. The coefficients of collaborative representation of test samples are sparse and robust to occlusion or other disguises. It is known that in the recognition or classification tasks, the discriminative model is also very important. The proposed model has two-fold advantages. It can represent the test sample well using redundant representation with sparsity and robust to disguises. On the other hand, the representation coefficients are generated with more discriminative information. It is very helpful for visual recognition issues. The point is that the l2 norm can achieve comparable performance to the l1 norm with simple implementation. Experimental evaluations on some benchmarks indicate that the proposed method could achieve impressive performances in terms of accuracy and efficiency with other existing works.