Facial Feature Extraction And Recognition Based On Dictionary Learning

Researches on face recognition have been continued for decades.Numerous researches have emerged in order to solve the problem that face recognition performance is affected by various factors such as lighting,pose,occlusion and lack of training samples.Meanwhile,high-dimensional face images make face recognition tasks more difficult,because distances between features are closer which is difficult for classification in high-dimensional Euclidean space.Therefore,feature extraction and classification is key to face recognition.Whether features can effectively describe facial information will ultimately determine the outcome of the classification.How to represent a face effectively through a set of complete features has always been pursued by academia.The space that complete feature set lies should include all facial information.In addition,the feature that can fully represent the object or facial information should not only be complete,but also should be well explainable.Therefore,this paper mainly studies and seeks a technical method to achieve the extraction of complete facial features.Sparse Representation（SR）has long received attention in computer vision since2006.The main idea is to encode the signal through a task-specific（and usually over-complete）dictionary under the constraints of the l₀ norm,and to suppress the Laplacian noise through sparsity constraints.Dictionary learning method was originally based on the research of sparse representation.It seeks to learn a set of redundant and complete features to form a dictionary for signal reconstruction.Later,dictionary learning is extended to realize the classification task by introducing a supervised term on the dictionary representation coefficients.Similarly,such a set of redundant,complete dictionaries can be used to represent facial information.When a set of complete feature dictionaries for face recognition tasks is extracted,it also has a clear explanation that different dictionary atoms correspond to different categories or attributes.However,traditional dictionary learning algorithms rely too much on pre-extracted facial features.The quality of input features limits the performance of dictionary learning algorithms.Regarding the issue above,we can establish a unified optimization objective for feature learning and dictionary learning and perform facial feature extraction and dictionary learning simultaneously.Therefore,the dictionary learning method is used to learn a complete face feature in this paper.The completeness of this feature depends on the property of dictionary learning algorithm itself.Specifically,the main contributions of this thesis are summarized as follows:1)Dictionary learning algorithm based on sparse embedding is proposedA Sparse Embedded Dictionary Learning（SEDL）algorithm is proposed to over-come the shortcomings of constraints on sparse representation coefficient in existing dictionary learning algorithms and the disadvantages of feature extraction.The algorithm maps high-dimensional face images into low-dimensional linear subspace and introduces the concept of metric learning by adding supervised constraints on sparse representation coefficients.It can finally learn a set of dictionaries that can represent face category information.Meanwhile,in the procedure of mapping features from high dimension to low dimension,orthogonal constraints are added to obtain a set of basis features with orthogonality.The orthogonality guarantees high fidelity in the reconstruction of the original signal and reduces the loss of information caused by dimensionality reduction.The algorithm combines the linear feature extraction based on orthogonal linear subspace projection with the improved supervised dictionary learning algorithm to establish a complete model and cost function,so that the dimensionality reduction and dictionary learning have a unified goal of reducing the loss of effective face category information during the process of dimensionality reduction and ultimately improve the accuracy of face recognition.2)Shared dictionary learning algorithm based on sparse embedding and non-linear feature extraction method is proposedMost studies on dictionary learning are limited to linear feature extraction and the study of the dictionary structure remains to be further.Therefore,this chapter focuses on the aspects of nonlinear feature extraction and dictionary structure,and expands the previous work.Firstly,the influence of dictionary structure on face feature extraction and recognition is considered.The concept of shared dictionary is introduced into the algorithm and a Shared Dictionary Learning algorithm with Sparse Embedding（SDLSE）is proposed.Secondly,from the perspective of feature extraction,nonlinear feature extraction is used instead of orthogonal linear feature extraction and then non-linear feature extraction based on shared dictionary is studied.SDLSE is extended to nonlinear feature extraction and a Nonlinear Shared Dictionary algorithm with Sparse Embedding（NSDLSE）is proposed.The effect of dictionary structure on facial feature extraction and recognition is studied on purpose and the concept of shared dictionary is introduced into the algorithm.Through the partition of the dictionary structure,the shared dictionary contains non class-specific information such as expression,lighting and pose,so as to achieve the extraction of complete face features.Secondly,based on the shared dictionary,the nonlinear feature extraction is introduced to improve the defect that the performance of dictionary learning algorithm depends on the original input features.3)A new deep learning layer,namely a structured layer in dictionary learning,is proposedAlthough the optimization problem above can be solved by alternating iterative method which can make dictionary learning algorithm get rid of the dependence on the original features,the alternate iterative method still confronts the problem of parameter initialization.Furthermore,deep convolutional network is only used as a feature extraction method whose good property of end-to-end is ignored.Therefore,deep convolutional network and dictionary learning algorithm are combined to establish a unified model and a deep learning network layer named as Dict Layer is proposed in this paper.Deep learning algorithm has the property of end-to-end characteristic that can achieve feature extraction and classification simultaneously.But it lacks a clear explanation of the meaning of its network layer.Otherwise,dictionary learning algorithm has a clear explanation.The supervised structured dictionary can effectively explain the meaning of the dictionary representation coefficients.Only the coefficients belonging to the same category are activated,but the feature extraction before limits the final performance.Therefore,those two methods are combined together in this paper.The correlation and the consistency of mathematical expressions between them are analyzed.The mature technology of structured dictionary in dictionary learning algorithm is introduced into Fully Connecting Layer（FC Layer）in order to achieve a unified model and cost function.The network layer can be considered as an improvement of the existing FC Layer and it can be trained through traditional Back Propagation（BP）,which avoids problems such as parameter initialization caused by traditional alternative iterative algorithm.The algorithm proposed combines the nonlinear feature extraction based on deep convolutional network with the structured information in dictionary learning to establish a unified model.This research also provides a feasible plan for introducing mature improvements from dictionary algorithms to deep learning.