Font Size: a A A

Research On Minimum Squared Errorbased Face Feature Extraction And Classification Algorithms

Posted on:2015-04-10Degree:DoctorType:Dissertation
Country:ChinaCandidate:Q ZhuFull Text:PDF
GTID:1108330422992558Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
With the booming development of computer and internet, the pattern recognition technology needs to deal with increasingly complicated data. One of the key tasks of pattern recognition is to develop the effective feature extraction and classification methods. Minimum squared error analysis based methods play very important roles in pattern recognition. For example, sparse representation based classification method (SRC) and the two-dimensional principal component analysis methods (2DPCA) are in the scope of minimum squared error analysis based methods. Based on the existing works, this dissertation proposes a number of minimum squared error based feature extraction and classification methods, and applies them into face recognition problem. The main works and the innovations of this dissertation include:We propose the local minimum squared error based classification method (MSEC). The traditional MSEC uses the unique sample label prediction model for all the test samples. So it performs not well in some complicated classification problems. The local MSEC proposed in this dissertation determines the nearest neighbors of the test sample, and uses them to construct the MSEC model for it. This method makes the test sample participate in the learning procedure of its label prediction model, which is beneficial to improve classification accuracy of the test sample.We propose the multi-subspace and multi-directional2DPCA (MS2DPCA and MD2DPCA). MS2DPCA directly works on the matrix, and does not need to transform matrix into high-dimensional vector, which can preserve the structural information embedding in the matrix. MS2DPCA is more capable in data representation and reconstruction than2DPCA. The experiment results on face databases demonstrate MS2DPCA with Gabor features achieves higher accuracy than previous methods. MD2DPCA allows performing two-dimensional feature extraction in any direction of the image matrix, whereas the traditional two-dimensional methods only can be used in vertical or horizontal direction. Hence, MD2DPCA is more general than2DPCA. By fusing the feature extraction results obtained by MD2DPCA, we design the face recognition algorithm that can take full use of discriminant information in these features.In order to improve the linear representation based classification model, we propose the nonlinear sparse representation based classification model. All the previous linear representation classification methods are based on the assumption that the test sample can be represented as the linear combination of the training samples. But this assumption does not hold in many face recognition problems. The nonlinear sparse representation based classification method assumes that the test sample can be represented as the nonlinear combination of the training samples. By introducing the kernel function, we solve the spare representation and classification problem in high dimensional feature space.We analyze the advantages and disadvantages of the linear representation based classification models with l1norm and l2norm optimization, respectively, and propose the fusion method that can take the advantages of the two optimizations. The experimental results on face databases demonstrate that the fusion method is superior to either l1norm or l2norm optimization based method.Based on the study of the feature extraction problem of linear representation based classification method, we propose the class coding error discriminant analysis (CCEDA) and fast sparse representation based classification method (FSRC). CCEDA aims to transform the sample into a new space, in which the sample has the maximum ratio between intra-class representation error and inner-class representation error. As a result, the sample can be more easily classified in the CCEDA space. The face recognition experiments show that the accuracy of the linear representation based method increases greatly after introducing the CCEDA transform. FSRC is based on the finding that the sparse representation model in PCA space is similar to that in the original space. In this method, we solve the sparse representation model in low dimensional PCA space and compute the class specific representation error in the original space. This method gets the approximation class specific representation with very low computation cost. The experiment results show that compared to SRC, FSRC has much higher classification efficiency and comparable classification accuracy.In summary, with the aim of improving the classification accuracy and classification efficiency, this dissertation proposes a number of minimum squared error based feature extraction and classification methods with application to face recognition. The experimental results show that the proposed methods achieve the desirable performance, and have wide application prospects in face recognition.
Keywords/Search Tags:minimum squared error, two-dimentional principal component analysis, feature extraction, sparse representation, face recognition
PDF Full Text Request
Related items