Feature Extraction And Classification Algorithm And Their Application In Face Recognition

Owing to the merits of being natural,directly perceived,convenient and the extensive application foreground,face recognition is one of the most important research topic in pattern recognition and artificial intelligent.During the past few decades,remarkable advances have been made in face recognition.Especially in the past ten years,a lot of theories and algorithms have been proposed,which is inspired by the development of physiology and cognitive sciences in face recognition.During the two-dimensional face image acquisition,a lot of discriminant information will be lost.Also,the varieties of illuminations,facial expressions and postures cause the face recognition to be a great challenge.Face recognition based on two-dimentional images is a classic small sample size problem.One of the greatest challenges of representation based face recognition is that the training samples are usually insufficient.Our work is focusing on the effect of insufficient training samples on the face image recognition.We try to improve the robuseness of face recognition on feature extraction and classifier designment.The main contribution in this dissertation can be summarized as following.(1)A novel feature extraction method induced from classifier is proposed.The characteristic of CRC is fully considered and a novel feature extraction method which is induced from CRC is proposed.Firstly,we get the coefficients of all face samples by collaborative representation.Then we define the inter-class reconstructive errors and intra-class reconstructive errors for each sample.After that,Fisher criterion is used to get the discriminative feature.At last,CRC is executed to get the identification results in the new feature space.So the feature space we get fits the classifier better.Experiment results on several face databases show that the proposed method is more effective than other state-of-the-art face recognition methods.(2)A Penalized Collaborative Representation based Classification(PCRC)for face recognition is proposed,which penalizes the "far" training sample to improve the performance of recognition.PCRC first uses the original collaborative representation to compute the distance between each training and test sample,and then treats these distances as penalized coefficients to design the penalized collaborative representation.The experiment results on multiple face databases show that our classifier,designed according to PCR,has a very satisfactory classification performance.Also,we implement PCRC on the features that are gotten by Gabor filters to improve the performance and robustness of PCRC.(3)Two improvement methods based on virtual training samples are proposed to improve the robustness of face recognition.Generating virtual training is an effective method to enlarge the training set.We think about that the virtual training set conveys some reasonable and possible variations of the original training samples.In order to improve the robustness of face recognition methods based on virtual training samples,we proposed two improvement algorithms.First,a generation virtual training samples that induces from collaborative representation based on classification is proposed.The proposed method calculates virtual training samples based on a reasonable objective function.The method allows virtual training samples gotten from our method to be little far from the original training samples.So they are able to contain proper variations of original face image such as variations caused by changeable illuminations,facial postures and expressions.Second,we design a new object function to get the representation coefficients generated from the original and virtual training set similar.In order to further improve the robustness,we implement the corresponding representation based face recognition in the kernel space.It is noteworthy that any kind of virtual training samples can be used in our method.We use noised face images to obtain virtual face samples.Experiment results on several face databases show that the two proposed methods are effective.(4)Two improvement methods based on MSEC are proposed to improve the robustness of face recognition.Classification is also an important step on face recognition.The fine classifier can get high robustness while the face images include all kinds of varieties.In order to improve the robustness of MSEC,two methods are proposed in this paper.On the first method,we impose a new projection matrix on the MSEC model,so the proposed method can tolerate varieties of samples.This can overcome the following drawback of MSEC:as the only one projection matrix is exploited to map training samples to the corresponding labels,the obtained mapping seems not to be very powerful and the algorithm is not very competent.On the second one,Kernel Iterative Minimum Squared Error Classification(KIMSEC)is proposed.The proposed classification design a new objective function to search a new projection matrix whose projection results not only approximate to the original class label but also to the approximation of class label.By this way,the new projection matrix can suit some potential variations of training samples or test samples.For further improvement of robustness,we extend it to the kernel space.