Research On The Improvement And Expansion Of Collaborative Representation Classifiers For Face Recognition

As one of the most convenient and effective biometric recognition systems,face recognition has always been a research hotspot in the field of pattern recognition.As its representative,the collaborative representation classifier has attracted more and more attention from researchers.In the field of face recognition,although collaborative representation classifier has the same performance as sparse representation classifier and has lower computational complexity,the classification mechanism of collaborative representation is still unclear.At the same time,face images in face recognition have the characteristics of high dimensionality and large amount of data.Therefore,how to extract effective discriminant features from high-dimensional images in order to recognize target images from a large number of face images quickly is a key point of face recognition.In addition,because image set contains more individual information than single image,face recognition based on image set can achieve better performance than face recognition based on single image.However,it also intensifies the urgency of finding efficient and effective classifiers and feature extraction methods matching the classifier.However,the current popular feature extraction algorithms,such as principal component analysis,are independent of specific classifiers.Therefore,this paper will reinterpret the collaborative representation classifier in different applications from the perspective of probability.At the same time,a matching feature extraction method is designed by using the new classifier's classification criterion based on minimum reconstruction residuals.Experiments on different databases show that our methods performs better in classification accuracy and stability than the contrast method.The main work of this paper is as follows:(1)To analyze the classification mechanism of collaborative representation classifier,the image based collaborative representation classifier is explained from the perspective of probability.Combined with classifier classification criteria,an orthogonal discriminant projection method based on probabilistic collaborative representation classifier is proposed.The recognition ability of the classifier is improved by obtaining the low dimensional features which match to the probability collaborative representation classifier.(2)The idea of probabilistic collaborative representation based on image is applied to image set based classification,and a probabilistic collaborative representation classifier based on image set is proposed.The orthogonal discriminant projection method is designed based on the classification criteria of the classifier and the characteristics of the image set.The low-dimensional features suitable for the classifier are obtained and the experimental results show that the proposed method can improve the classification performance of the classifier.(3)A cost sensitive probabilistic collaborative representation classifier based on image set is proposed.The basic idea of the classifier is to introduce cost information in the stage of obtaining discriminant projection matrix and the stage of classifier design,so that both the obtained projection matrix and classifier have cost-sensitive functions.Therefore,a lower false recognition rate can be obtained at the classification stage.Experiments show that our method can achieve a lower false recognition rate while no significant improvement in rejection rate.