New Methods Of Representation-based Classification Learning And Its Application In Face Recognition

Recently,the theory of representation-based classification learning as one of the most important research topics have obtained notable research achievements in the field of face recognition with an innovative perspective.However,there are still some fundamental constraints remain unsolved.Firstly,the features of training images are not fully embodied and utilized,leading to the uncertainty of final results.Secondly,the sufficient images are necessary for robust recognition.Thirdly,the time complexity of representation-based classification methods is so high that it can't meet the efficiency requirement of practical application,etc.To deal with the above problems,we propose five improved algorithms in this paper with the deep study on the theory of representation-based classification learning.The specific work is as follows:(1)We analyze several typical and representative representation-based classification methods from three aspects.Firstly,we conduct the research on the technologies of feature extraction in which Principal component Analysis(PCA),Linear Discriminant Analysis(LDA)and Convolutional Neural Network(CNN)are summarized;Secondly,the algorithms about expanding the training set including Extend SRC(ESRC)and Integrate the Original Face Image and its Mirror Image for Face Recognition(IOMFR)are also studied;Thirdly,we explore some methods with improved classification strategies,such as Two-Phase Test Sample Sparse Representation(TPTSR),Sparse Representation-based Classification Method using Iterative Class Elimination(SRICE),as well as Conventional and Inverse Representation-based Linear Regression Classification(CIRLRC).Finally,we can be inspired by learning the process and rationales of these algorithms which provide foundations for the future work.(2)Focusing on the problems of low accuracy and high computational complexity of conventional representation-based classification methods in small sample size face recognition,we propose four fast SRC methods using quadratic optimization in downsized coefficient solution subspace based on ESRC:extended sparse PCA(ES-PCA),extended PCA based ESRC(EP-SRC),extended sparse LDA(ES-LDA)and ES-PCA using CNN-based features(ES-PCA-CNN).The four improved methods apply a set of dimensionality reduction constraint models to achieve a compressive linear representation of the test sample.The solution of this optimization using l1-united with l2-united minimization can successively be built.By designing.the more accurate reconstruction of the test sample using feature coefficients can be achieved for robust classification.Then,this paper analyzes the correlations,advantages,disadvantages,and time complexity of the proposed methods,respectively.Finally,experimental results conducted on several well-known face datasets demonstrate the merits of the proposed methods.(3)We propose a bi-directional collaborative representation-based classification algorithm via convolutional neural network(BCRC-CNN)based on CIRLRC in consideration of the problems of over-fitting and data uncertainty in face recognition.More specifically,the proposed algorithm firstly employs the pre-trained VGGNet-16 model for feature extraction to alleviate the adverse information from original dataset,and then develops an efficient bi-directional representation model to obtain the corresponding coefficients of samples by integrating the collaborative information from both training samples and test sample.The last contribution of the proposed method is to utilize a competitive fusion method weighting the reconstructed residuals from the bi-directional representation model for robust face recognition.Experimental results obtained from a set of well-known face databases including AR,FERET,ORL,LFW and FRGC verify that the effectiveness of the proposed method,whether there are sufficient training samples or not.