Research On Face Recognition Based On Multi-scale Patch Collaborative Representation

Small sample training is one of the most challenging problems in face recognition.Multi-scale patch collaborative representation of face recognition method uses majority voting method to determine the label of a sample.However,at a scale,all patch blocks formed by a sample often belong to many different categories.Using the majority voting method to make decisions ignores the possibility that the forecast samples belong to other categories.Therefore,this paper proposes a multi-scale patch face recognition collaborative representation fusion method based on fuzzy decision.In a scale,among all patch blocks divided into a sample,the degree to which a sample belongs to a category is expressed by the proportion of the number of patch blocks belonging to a certain category to the total number of patch blocks.In this way,at each scale,the sample set will get a fuzzy decision matrix,and each element of the fuzzy decision matrix represents the possibility that each sample belongs to each category,thus solving the absolute problem of classification.Then the fuzzy decision matrix is multiplied by the real label,and the fuzzy decision accuracy of the sample set at this scale will be obtained.Different weights are applied to the fuzzy decision accuracy obtained at different scales,and the integration of multi-scale outputs is realized through regularization boundary distribution optimization.A large number of experiments show that this method has higher recognition accuracy and is superior to many patch-based face recognition algorithms.In the application of face recognition,the number of samples corresponding to different image categories is different,which will have different effects on the fusion weight of training.Based on this problem,this paper puts forward the concept of category weight matrix based on information fusion to realize the influence of different categories on image recognition due to the different number of samples.In the process of information fusion,the traditional multi-scale patch collaborative representation method only considers the classification information at different scales,and ignores the influence of different samples on the classification results.Each row vector in the weight matrix represents the corresponding weight value of different categories at a certain scale,and each element of the weight matrix is the weight value of a certain category at a certain scale.Then,the identification accuracy of the sample set belonging to this category can be obtained by multiplying the decision value of the sample set belonging to a certain category in each scale with the weight vector of the corresponding category in the weight matrix.Then,the loss function is constructed and optimized.A large number of experiments show that this method has a high recognition accuracy,which is better than many patch-based face recognition algorithms.