Consistent Sparse Representation And Its Application Based On Image Set

Face recognition is one of the hot topics in the field of computer vision.The research of face recognition is still challenging due to the serious occlusion,illumination changes and complex background.In recent years,face recognition algorithms based on image set is getting more and more attention.Because every video sequence are accompanied by those significant change.It is better able to deal with the challenge.In this paper,we propose an efficient algorithm named Consistent Sparse Representation.First,we obtain a prior that images belong to the same class should be treated as a whole.So images from the same test sequence can be considered as a whole.It can be represented by one or several classes of sample data set.We should consider the group-sparse of sample data set and test set simultaneously.The property is defined as label consistency.Second,we propose a new mixed norm F,0l to describe recovery coefficient.The purpose is that test set can be represented by images from one class in sample data ideally.As a consequence,the recovery coefficient has block structure.We can determine the category of test set according to the label of non-zero elements.Then,we propose a novel algorithm based on ADM to solve recovery coefficient because F,0l norm is discontinuous.The F,0l norm can be approximated to 0l norm by introducing new variables to extend the original items.Finally,we compare six algorithms on several challenging databases,such as COX,Honda and so on.Experiments demonstrate that CSR algorithm is more effective than other algorithms.CSR algorithm can be more robust to occlusion,posture changes.In addition,the algorithm is applied to tracking framework.For each candidate image,we create video stream and construct non-object basis.The tracking problem is transformed into two classification.CSR algorithm is compared with other four algorithms on the open test sequences,the results show that the validity of our algorithm.