Research On Dictionary Learning In Image Classification And Application

Image classification is a fundamental and important issue in computer vision and pattern recognition.The research of image classification can accelerate the development on network im-age retrieval,intelligent video surveillance and image scene understanding,etc.The accuracy of image classification system not only depends on the extracted feature,but also the classifier adopted in the system.Sparse representation-based classification?SRC?has been recently pro-posed and applied in robust face recognition.Since the SRC scheme achieves competitive perfor-mance in face recognition?FR?,it attracts the researchers'interest in sparse representation-based pattern recognition.With the understanding of SRC in depth,researchers have been aware that the dictionary adopted in SRC played key role.This thesis concentrates on dictionary learning based image classification.In order to im-prove the classification accuracy of the image classification system through dictionary learning method,not only discriminative ability of the learned dictionary should be boosted,but also the technologies in pattern recognition,such as dimensionality reduction and classifier,should be unified into the dictionary learning framework.On the other hand,reducing the computa-tion cost of dictionary learning is an urgent problem.The main innovations of this thesis are summarised as follows:?1?This paper proposes a supervised class-specific dictionary learning?SCSDL?model based on sparse representation to learn a structured dictionary to perform classification.In the SCSDL model,besides the representation-constrained term,the coefficients diversity term is further proposed to enhance the discriminative ability of the learned dictionary.When perform classification,because of the introduced coefficients diversity term in SCSDL,both the represen-tation residual and the representation coefficients are discriminative,thus a novel classification scheme associated with SCSDL is presented to exploit such discriminative information.Exper-iments validate that the proposed framework obtains competitive performance compared with the state-of-the-art methods.?2?Based on the proposed SCSDL model,this paper proposes a simultaneous dimensional-ity reduction and dictionary learning?SDRDL?model by introducing the projection-constrained term,to learn a DR projection matrix and a class-specific dictionary?i.e.,the dictionary atoms correspond to the class labels?simultaneously.Most existing methods perform dimensional-ity reduction?DR?and dictionary learning?DL?independently,which may result in not fully exploiting the discriminative information of the training data.Since simultaneously learning makes the learned projection and dictionary fit better with each other,more effective pattern classification can be achieved using the representation residual.Experimental results on a series of benchmark image databases show that our proposed SDRDL outperforms many state-of-the-art discriminative dictionary learning methods.?3?Since the computation cost of SDRDL is high,this paper proposes discriminative di-mensionality reduction projection and dictionary pair learning?DDRPDPL?framework based on the collaborative representation to reduce the computation cost.Despite SRC's initial suc-cess in face recognition,some research works reveal that it is not the sparse representation but the adoption of collaborative representations in general plays a more crucial role in the suc-cess of the SRC,and then proposed the collaborative representation based classification?CRC?,where the l₂-norm is used to regularize the coding coefficients.Compared with SRC,CRC could obtain the analytic solution for the coefficients with the low computation cost.Therefore,when DDRPDPL updates the dictionary,the coefficients can be solved by the analytic solution with efficient computation because of the l₂-norm.The extensive experiments on image classifica-tion tasks such as face recognition and objection classificationre showed that DDRPDPL could achieve similar classification accuracy to SDRDL achieves,but the computation cost is much reduced.?4?This paper proposes a SVM multi-class loss driven dictionary learning?SMLDDL?model.SMLDDL learns a dictionary while training a SVMs over the representation coefficients,then SVMs drive the dictionary by the designed SVMs multi-class loss function.Therefore the SVMs and dictionary can be obtained together,thus they are fit better with each other for performing image clssification.To perform action classification in video by using SMLDDL,this paper further proposes motion improved WLD?MotIWLD?to represent the action in video.By ultilizing MotIWLD,SMLDDL learns a dictionary and a SVMs to perform classification on a query sample,which is represented by MotIWLD.Experimental results validated that our proposed methods are effective and discriminative.