Research On New Methods For Deep Collaborative Representation-based Classification And Its Applications

Face recognition is one of the challenging topics in the field of pattern recognition and artificial intelligence.It is widely used in information security,financial security and public security,which has produced enormous economic and social benefits.However,due to the high data dimension of the original face image and the interference such as noise,illumination,occlusion and pose,it is impossible to directly match and recognize the original face image.Therefore,the quality of face representation directly affects the accuracy of subsequent classification.How to effectively represent the original images and extract appropriate features for rational use is a core problem in the field of image classification.As a new feature representation method,sparse representation can effectively solve the existing in the practical application of the information redundancy issue such as difference of large,high computational complexity and interpretability.It is widely used in face recognition in recent years,becomes a research hotspot in the field of image classification.A large number of sparse representationbased and collaborative representation-based face recognition framework is put forward.At the same time,the emergence of deep learning theory and the rapid development of its algorithm demonstrate its excellent characteristics of using a large amount of data to describe image features,which makes the deep learning model based on feature extraction become a research hotspot in the field of image recognition and classification.Therefore,this paper will take face recognition as an example to further research the collaborative representation-based and deep learning-based image classification algorithms.It is aimed at many internal and external factors(such as noise,illumination,occlusion and pose)that affect face imaging in unrestricted environment.So as to obtain efficient image feature representation and improve the system classification performance.Aiming at the above problems,this paper deeply studies and analyzes the sparse representation-based face classification method and the related improved algorithm.The paper summarizes the principle and process of some classical algorithms from two aspects: classification strategy and sample set expansion.In addition,special attention is paid to the research and application of deep convolutional neural network and 3D face reconstruction method-based in face classification.On this basis,this paper proposes three improved collaborative presentation-based algorithms for face classification,the specific contents are as follows:(1)From the perspective of improving classification strategy and recognition efficiency,this paper presents a bi-directional CRC algorithm using convolutional neural network-based features for face classification(BCRC-CNN).Firstly,employ a deep convolutional neural network to extract facial features from the original gallery and query sets,and then develop a fast reverse representation model to obtain the auxiliary residual information between each training sample and the reconstructed one that is achieved from the test sample.Secondly,offer a new solution to the bi-directional optimisation problem by which the input sample is well represented by the forward linear combination and the reverse one,respectively.The last contribution is to utilise a competitive fusion method for robust face recognition,which weighted reconstructed residuals from the bi-directional representation model.Experimental results obtained from a set of well-known face databases including AR,FERET,and ORL verify the validity of the proposed method,especially in the robustness to small sample size problem.(2)From the perspective of improving classification strategy and identification accuracy,this paper presents a discriminative bi-directional CRC algorithm for image recognition(DBCRC).Firstly,a structured dictionary was obtained using discriminant dictionary learning(FDDL)model,in which the sub-dictionary of specific-class has robust representation capability for the samples of related-classes.Therefore,by means of the larger inter-class and the smaller intra-class scatters,the reconstruction error and coding coefficient can be described in discriminant manner.Then,utilize the obtained sparse coding coefficients as the test samples for the bi-directional representation,constructing a fast inverse representation model.The bidirectional reconstruction based residual information between each test sample and the structured dictionary is estimated via bi-directional representation strategy.Finally,the competitive fusion method is used to achieve the final classification result,by weighting and ranking the obtained reconstructed residuals from the bi-directional representation model.Experimental results conducted on a set of well-known face databases including AR,CMU-PIE and LFW verify the effectiveness of this algorithm,especially in the robustness to the problem of occlusion in training samples.(3)From the perspective of improving classification strategy and recognition efficiency,this paper presents an extended CRC algorithm using 3D morphable face models(3D-ECRC).Firstly,the 3DMM model was fitted with the 2D face image in the dictionary to reconstruct the 3D shape and texture of each image,and a series of virtual 2D face images with attitude changes were rendered to enrich the attitude information in the database.Then,the original training sample and the virtual sample are combined to build an extended dictionary,the features are extracted by using the deep convolutional neural network,and the reconstruction process of the extended dictionary and test sample is mapped to the CNN feature space.Finally,collaborative representation-based classification is used to complete the classification.The experimental results on the equal public face data set including FERET and LFW prove the effectiveness of the proposed method,especially the robustness of the problem of pose changes.