Research Of Algorithm Based On Consistent Sparse Representation For Face Recognition

Face recognition is a popular research topic in computer vision and pattern recognition.With inspiration from the sparsity mechanism of the human vision system and the success of sparse coding in image processing,the sparse representation based classification algorithm has received sufficient attention and achieved excellent performance in face recognition.With the elevation of the importance of robust representation in face recognition,the regression analysis based classification methods are showing many advantages.Particularly,l1-norm based sparse representation classifier builds a constraint residual to achieve robust representation against noise,which assumes that errors of pixels follows sparse distribution.But in some cases,contiguous occlusion and real disguise still affect recognition accuracy greatly,because it is obviously problematic that the occlusions are coded by clean training samples.On the other hand,it is impossible to pre-design a comprehensive occlusion dictionary covering large amount of noise atoms.In Sparse representation,a test image is encoded by a sparse linear combination of training samples.The L1-regularizer used in sparse representation is beneficial to produce a good reconstruction of the test face image with sparse error,but it is incapable to guarantee the robustness against local structural noise.To tackle this issue,we propose two methods on robust face recognition.The highlights and main contributions of this paper include:1.We propose a novel method namely consistent sparse representation,which is based on the idea that only undisturbed blocks/patches of images have similar coding coefficient vectors.In consistent sparse representation,a multi-coding sparse representation model is represented and in this model consistency of coding vectors and fidelity of representation are balanced.We consider sparsity and consistency simultaneously in optimization for obtaining significant sparse coefficients with better effect of classification.Revealing experiments in face recognition demonstrate the robustness of our methods to random block occlusion,local illumination and real partial disguise.2.To enhance noise tolerance of sparse representation based classifier,we propose an improved L1-regularizer based on trimmed sparse coding by using an extra penalty on correlation among all coding coefficients.Different from traditional single-coding scheme in sparse representation,we use multiple coding coefficients to represent patches of a test image by its corresponding training patches.The consistency penalty imposed into the new sparse representation model improves the confidence for accuracy classification.Experimental results show the superiority of trimmed sparse coding on two benchmark databases,and it outperforms other state-of-the-art methods.