Research On The Technologies Based On The Sparse Representation Model And Discrimination Anylasis For Face Classification

The technology of face recognition is an important branch in the field of biometric recognition,because the high dimensional data information of human faces is sufficient to support effective personal identity verification,which attracts a lot of researchers' interest both due to its challenging nature and due to its diverse set of applications.Therefore,in recent years,the research and application of face information emerge in an endless stream.And a series of different feature extraction and classification algorithms have been proposed and achieved good performance.In our work,under the framework of LDA(Linear Discriminant Analysis)and SRC(Sparse Representation-based Classification),structure information is incorporated in the process of feature extraction method and more discriminant information is preserved in the extracted feature space.Nevertheless,most of the SRC methods use the original training samples as the dictionary.The nonzero coefficients are expected to concentrate on the training samples with the same identity as the test sample and on the synthesized samples.In real-world applications,the face images are captured under different condition of poses and illuminations.The side-effect of illumination can somewhat be overcame by the procedure of construct samples.So an important issue that whether an optimal dictionary can be derived from training data needs to be further discussed.Very recently some researchers have started to question the use of sparsity in image classification,which revealed that it is the collaborative representation(CR)mechanism,but not the l1-norm sparsity constraint,that truly improves the face recognition(FR)accuracy.This interesting conclusion puzzled us why the relaxed sparsity constraint achieved better result.Recently,we have investigated the construction of the dictionary and training variables selection strategy for test sample based sparse representation.It deserves to be mentioned that a strict and standard dictionary will create sparser coefficient,but it might not achieve better classification accuracy in SRC model.Therefore,constructing a rational dictionary and training variables selection strategy for SRC model is a challenging task for us.In the study of traditional F-LDA,the algorithm combined with FKNN is considered to regain the statistical properties of the patterns such as mean value and scatter covariance matrices.However,after:investigating the membership allocation formula,we find that the F-LDA method attempts to fuzzify or refine the membership grades of the labeled patterns only by fuzzifying each class center.How can we take full advantage of the distribution information of each sample to the redefinition of scatter matrices?In our work,we extended the F-LDA and included complete fuzziness in the calculation of between-class scatter matrix and within-class scatter matrix.The main achievements and innovations are listed as follows:(1)We propose a novel SRC fusion method using hierarchical multi-scale Local Binary Patterns(LBP)and class-based elimination strategy for face recognition.The proposed method involves three aspects:dictionary optimization,training variables selection and classification strategy,sparse coding coefficient decomposing.Actually,a strict and standard dictionary will create sparser coefficient,but it might not achieve better classification accuracy in SRC model.Therefore,constructing a rational and optimal dictionary for SRC model is a challenging task for us.Understanding the good performance of such unconventional dictionaries demands new algorithmic and analytical techniques.First of all,we get an optimized dictionary through extracting hierarchical multi-scale LBP features from the original training samples.Second,the training variables selection and classification strategy aims to represent a query sample as a linear combination of the most informative training samples,and exploits an optimal representation of training samples from the classes with major relevant contributions.Instead of eliminating several classes at one time,we choose eliminating classes one by one with greedy search(GS)sparse coding process until the predefined termination condition is satisfied.The final remaining training samples are used to produce a best representation of the test sample and to perform classification.In the context of the proposed method,an important goal is to select a subset of variables that accomplishes one objective:the provision of a descriptive representation for sparse category knowledge structure.Experimental results which are conducted on the ORL,FERET and AR face databases have demonstrated the effectiveness of the proposed method,and the recognition rates are increased at least 2 percent compared with other representation-based methods.(2)We develop a framework that fuses virtual synthesized training samples as bases and sample-based elimination strategies for classification.In general,images of faces are not strictly captured under a frontal and natural pose.More specifically,in the first stage of the proposed method,sampling uncertainty of the linear approximation model is alleviated effectively by constructing extra synthesized mirror training samples which generated from original images.Subsequently,in the second stage,features are extracted from the above dictionary using a hierarchical multi-scale Local Binary Patterns(LBP)scheme.Once features are computed we use Principal Component Analysis(PCA)to project the features to a low-dimensional space.In the third stage,we combine the synthesized samples and original ones together to describe a test sample and simultaneously determine the relatively informative training samples by means of exploiting the reconstruction deviation of all the dictionary atoms.Experimental results which are conducted on the ORL,FERET,GT,Extend Yale B,LFW and AR face databases have demonstrated the effectiveness of the proposed method,and the recognition rates are excellent compared with other representation-based methods.(3)We introduce a generalized multi-scale Local Quantized Patterns(LQP)to represent the original image of the sample.The current local features descriptors,such as LBP,these local patterns use hard-wired codings and fixed layouts and are limited to very coarse quantization,mostly binary.These weaknesses limit the comprehensive information encoding capability of these local pattern features and prevent them from leveraging all the available information.But introducing too many neighborhoods will obviously increase the length of coding vector.Inspired by visual words method,the frequency of every code vector is counted firstly,and then training the off-line codebook use the k-means algorithm which satisfied the quantization conditions and as a lookup table.Using a lookup table,the corresponding quantized code of a given features vector is easily obtained.Here,we convert face image into high,middle and low resolutions,LQP features are extracted in each scale,which make full use of vector quantization and lookup table.(4)In this thesis,we propose a novel quaternion-based discriminant color space transformation to perform face recognition.The proposed method represents and classifies color images in a simple and mathematically tractable way,which is suitable for a large variety of real-world applications such as color image feature representation,extraction and classification.The proposed algorithm firstly uses quaternion number to denote the pixel in the color image and exploits a quaternion vector to represent the color image.Traditional methods focus on the gray-level image or convert the RGB color space to HIS space and deal with the intensity channel.(5)We develop a reformed fuzzy supervised learning algorithm.First,a reformative supervised fuzzy LDA algorithm(RF-LDA)for the training samples is proposed.Compared with the conventional fuzzy LDA algorithm,the presented algorithm computes the discriminant vectors associated with the membership grade from each training sample,which is theoretically effective in overcoming the classification limitation originating from the imprecise samples.Second,considering the fact that the Kernel Fisher Discriminant(KFD)is effective in extracting the nonlinear discriminative information of the feature space by using kernel trick,a kernel version of RF-LDA is presented subsequently,which has the potential to outperform the traditional fuzzy learning algorithms,especially in the cases of nonlinear small sample sizes.The advantage of this learning algorithm is that it successfully utilizes the improved kernel fuzzy LDA algorithm as a supervised feature extraction tool.Meanwhile,by means of the control parameter estimation,we address the problem that the particular value of offset in the calculation of the grade of membership is dynamically assigned.