Research On Feature Learning Based Face Recognition Under Unconstrained Scenario

In recent years,our country has invested large amounts of resources in the field of human-centered intelligent information processing and control technologies based on biological feature.Because of face recognition(FR)technology has many advantages,it has become the most popular technology in the field of biometrics recognition.Existing FR methods has been widely used in various constrained scenario.However,there are several issues haven't been solved yet in the unconstrained scenario.(1)Blurring face image issue in the unconstrained scenario:Face images which are captured by the mobile devices usually have low quality,motion blur,out of focus and low resolution.(2)Single extreme intraclass variation issue:In the unconstrained scenario,the intraclass variations(such as,expression,illumination,pose and so on)in face images usually result in variations extremalization.(3)Multiple intraclass variations appear simultaneously issue:In the unconstrained scenario,face images are vulnerable to various intraclass variations.In other words,there are various intraclass variations in a single face image.Based on extensive survey,the author found that existing FR methods haven't solved these problems well.Therefore,it is an urgent problem to be solved that how to improve the recognition/verification accuracy and the power of data-adaptive under the unconstrained scenario.The author studies the key and difficult problems of face recognition under the unconstrained scenario.The author focus on the pre-processing and feature extraction process which significantly affect the performance of face recognition methods under the unconstrained scenario.The main research works and contributions of this paper can be listed as follows:1.Proposing a weighted non-locally self-similarity sparse representation facial deblurring method to solve the blurring face image issue in the unconstrained scenario:Existing face deblurring methods either rely on large sample sets for estimating salient contour of blurred images,or rely on a large number of face images within the same class as reference images for face deblurring.These limitations lead to the existing face deblurring methods perform not well in the unconstrained scenario.In order to solve these limitations,the author proposes a Weighted Non-Locally Self-Similarity(WNLSS)face deblurring method.The author employs the inherent structural characteristic of face images and designs a facial deblurring method for face images,which does not require the salient contour estimation process or large numbers of reference images.The proposed WNLSS method significantly improves the performance of existing face deblurring methods.On the one hand,the WNLSS method employs sparse representation model to deblur the blurred facial image,so that the estimation process of salient contour can be removed.Therefore,this method solves the problem of the some existing deblurring methods which rely on the salient contour estimation process;On the other hand,the WNLSS method searches several non-local self-similar image patches for each image patch in the blurred facial image,and employs these non-local patches to estimate the sparse coding coefficients of the blurred image.Therefore,the proposed WNLSS method does not need to use any reference image for deblurring process,and solves the defect of the other existing deblurring methods which rely on the reference images.Moreover,the author also introduces the idea of weight coding into the proposed WNLSS method to further reduce the residual between the deblurred image and the original image.The WNLSS method makes the effect of those pixels with large residual is further reduced.Compared with other existing deblurring methods,our WNLSS significantly improves the performances of deblurring on both synthetic and real blurred face images.2.Proposing three kinds of learning-based binary coding face recognition methods to solve the single extreme intraclass variation issue:The author employs learning-based binary coding FR methods to solve the single extreme intraclass variations issue.However,these existing learning-based binary coding FR methods generally lack the feature descriptive power and the feature discriminative power.Moreover,it is easy for the existing binary coding FR methods to overfitting when faced with a small sample dataset.In order to solve these problems,the author proposes three kinds of independent yet closely related learning-based binary feature FR method.These methods are proposed to enhance the feature descriptive power,feature discriminative power and the ability of avoiding overfitting,respectively.(1)In this paper,the author proposes an Iterative Quantization Binary Codes(IQBC)face recognition method.The IQBC method combines the existing theory of multi-class spectral clustering and the orthogonal Procrustes analysis problem to learn a set of optimized projection matrices.These projection matrices rotate the mean-removed data into a linearly separable binary feature space,so that the quantization loss between real-valued features and binary features can be minimized.Moreover,in order to improve the descriptive power of our proposed method,the IQBC method requires the variance of each bit is maximized and the bits are pairwise uncorrelated.The author tests the performances of IQBC on four public face datasets(FERET,PEAL,LFW and PaSC).The results demonstrate that the IQBC method effectively enhances the descriptive power of the output binary features,and improves the performance of existing FR methods of the same kind.(2)In this paper,the author proposes a Spherical Hashing based Binary Codes(SHBC)face recognition method.The author intends to improve the discriminative power of binary features based on the following two aspects.On the one hand,the author employs the supervised CCA method to introduce the identity information into the SHBC method's training process,which enhance the discriminative power of SHBC binary features.On the other hand,the author employs the hyperspheres to define binary codes instead of hyperplane.The hypersphere-based SHBC method is more likely to assign the same binary code for two facial images with the same label,so the discriminative power of SHBC method can be improved.The author investigates the performances of SHBC on four public face datasets(FERET,PEAL,LFW and PaSC).The results demonstrate that the SHBC method effectively enhances the discriminative power of the output binary features,and improves the performance of existing FR methods of the same kind.(3)In this paper,the author proposes a Sparse Projection Matrix Binary Descriptors(SPMBD)face recognition method.The SPMBD method preserves the description power of binary feature as much as possible,and restricts the feature solution space by limiting the number of non-zero elements in the projection matrix,so that the complexity of SPMBD method can be reduced.The introduction of sparse constraint not only solves the overfitting problem,but also solves the problem of high time complexity which is caused by high-dimensional output features.The author investigates the performances of SPMBD on LFW and PaSC public face datasets.The results demonstrate that the SPMBD method effectively improves the data-adaptive power of the binary features,and further reduces the computational cost of the high-dimensional based methods.3.Proposing two kinds of subspace learning based deep learning face recognition methods to solve the multiple intraclass variations appear simultaneously issue:The author employs subspace learning based deep learning FR methods to solve the multiple intraclass variations appear simultaneously issue.However,there are following three problems in the existing deep learning methods,such as Convolutional Neural Network(CNN)and PCANet methods.These problems are the problem of training sample size,information bottlenecks problem,and single-scale pooling problem.In order to solve these three problems,the author proposes two subspace learning based deep learning methods.(1)In this paper,the author proposes a Spectral Regression Discriminant Analysis Network(SRDANet)face recognition method.SRDANet employs the supervised spectral regression discriminant analysis(SRDA)method to learn the eigenvectors of facial images and considers them as the convolutional(filter)kernel in the convolutional layer.The proposed SRDANet method provides a novel learning strategy for the convolutional kernels.This kind of convolutional kernel learning strategy ensures that the proposed SRDANet can use a small number of training samples to learn discriminative convolutional kernels through only one forward propagation.Therefore,the problem of training sample size can be solved.The author tests the performances of SRDANet on LFW face datasets.The results demonstrate that SRDANet can use less training samples to learn discriminative convolutional kernels.The SRDANet outperforms the performance of those existing CNN methods with the same size training samples.(2)In this paper,the author proposes a Multiple Scales combined Principle Component Analysis(MS-PCANet)deep learning FR method.The proposed MS-PCANet method cumulates the low-level feature and the high-level feature,and forms the final output feature which includes local detailed information and global holistic information.Therefore,MS-PCANet method effectively solves the information bottlenecks problem.Moreover,the author also introduces the idea of Spatial Pyramid Pooling(SPP)into MS-PCANet method.The SPP pools the feature map from multiple scales.So the output feature of pooling layer includes the local detailed information and the global structural information of feature maps.Therefore,single-scale pooling problem can be effectively solved.The author investigates the performances of MS-PCANet on LFW and PaSC public face datasets.The results demonstrate that MS-PCANet can effectively solve the information bottleneck problem and single-scale pooling problem,and improves the accuracy of deep learning based FR methods.