Group-based Multiply Face Images Joint Recognition Method

As a kind of biologic recognition method,face recognition is considered as one important field of computer vision research and industrial application for it is secure,not easy to lose,sampling-friendly and convenient.Due to good classification ability of Sparse Representation-based Classification(SRC),many modified sparse representation-based algorithms for face recognition are introduced.While sparse representation-based classification and related algorithms represent and classify each single sample one by one,it is of low efficiency in the case of knowing some multiply testing samples to be in the same class for not making use of the intrinsic testing-sample structure.Therefore,multiply face images joint recognition research is conducted in this paper with the knowledge that testing-sample group belongs to the same but label-unknown class :Based on SRC model,a novel robust method based on sparse representation-based classification with intraclass testing-sample group(IGSRC)is proposed,that is,firstly take multiple intraclass testing-sample into the same group,then adopt the matrix 1L norm regularized least squares classification,lastly classify the group by evaluating which class leads to the minimum representation error.Further,methods by substituting the class-wise coefficients norm for the representation error and before classification by introducing sparse concentration index(SCI)to select testing samples which are more favorable of classification to form new intraclass testing-sample groups are investigated.Meanwhile,by clustering the testing-sample in advance,intraclass testing-sample group could be obtained in the paper even if not knowing whether testing-samples are labeled as the same class or not.Numerical experiments in AR and Extended Yale B face database confirms the effectiveness of the proposed methods.SRC and IGSRC doesn't work well for small sample problem(especially single sample)while sparse and dense hybrid representation-supervised dictionary decomposition(SDR-SLR)algorithm offers a new approach to small sample problem by dictionary learning and sparse and dense hybrid representation.However,it still doesn't perform well when the number of training samples per class is as small as 1 or 2.Applying the idea of grouping interclass testing-sample,a new algorithm called IG-SDR-SLR is proposed.Numerical experiments show that it performs significantly better than SDR-SLR when the number of training samples per class 1 or 2 which validates the effectiveness of the proposed method.