Image Classification Based On Dictionary Learning

Image classification is a important branch of computer vision analysis,in which image representation plays a vital role.Sparse representation and low-rank repre-sentation are two typical representation learning techniques.Given an over-complete dictionary and a sample,sparse representation aims to find the sparest representation of the sample among all the linear combinations of dictionary atoms.Therefor,it could reveal the relationships of high-dimensional samples,and has been widely used in face recognition,image denoising and visual tracking.Sparse representation based clas-sifier(SRC)is a classical method in sparse representation,which achieves promising results in face recognition.Low-rank representation has shown excellent performance for learning features from noisy observations,since it is able to discover the underlying structures in noisy data.In both of sparse representation and low-rank representation,the dictionary D plays an important role and directly affects the discrimination of im-age representation.This dissertation starts from sparse representation and low-rank representation,focuses on dictionary learning based image classification to learn dis-criminative and robust image representations.Some creative or innovative works are presented as follows:A set of explanatory dictionary can help people better interpret data,and pro-vide new insights into underlying processes.To improve the explanatory of dictionary,traditional dictionary learning methods generally use sparse constraint or manually specified structural constraints.However,these methods are sensitive to noise and can not depict the complex structures among images adequately and accurately.In the study,Schatten-p(0<p<1)norm based principal component analysis(SpPCA)is proposed to solve the problems.In SpPCA,the dictionary atom is firstly trans-formed into a two-dimensional matrix form,then a low-rank constraint is utilized to characterize the structural information in atoms.Instead of using nuclear-norm to ap-proximate rank function,which may deviate the solution away from the real solution of original rank-minimization problem,the Schatten-p quasi-norm is used to approxi-mate rank function.The experimental results on image denoising and face recognition demonstrate the effectiveness and robustness of the proposed method.The performance of sparse representation based dictionary learning methods will deteriorate when the data is contaminated(i.e.occlusion,disguise,lighting variations,pixel corruption).Since low-rank representation based dictionary learning methods are robust to the noise among data,the classification performance can be significantly improved.In order to well cope with training samples with large noise,structured discriminative dictionary learning based on Schatten-p norm low rank representation(SDDL)is presented.Different from conventional low-rank representation based dic-tionary learning methods,which use low rank representation nuclear norm to approxi-mate rank function,SDDL makes use of Schatten-p(0<p<1)norm to approximate rank function.In addition,a weighted coefficient constraint item is added in the ob-jective function to increase the discrimination of dictionary,which encourages that the representation coefficient matrix have class-wise block-diagonal structure and the class-specific dictionary is able to well reconstruct the samples from the same class.Experiments on three databases verify the feasibility and robustness of SDDL.Among sparse representation based supervised dictionary learning methods,the dictionary size and the property of each dictionary atom are pre-defined.The dictio-nary size is generally set large to achieve good classification performance,which will cause big redundancy between atoms and encoding time increase.Moreover,the pre-defined property may not reflect the structure among data accurately.To solve this problem,a structure adaptive dictionary learning(SADL)method is proposed.The method is able to adaptively obtain the relationship between dictionaries atoms and class labels through learning a binary matrix,and abandon the atoms which are not relevant to any class.In addition,a discrimination term based on Fisher discrimination criterion is applied on the coefficients to enhance the discrimination of the dictionary,and a weighted dictionary coherence term is also added to reduce the correlation be-tween atoms.Experiments on different databases demonstrate the effectiveness of the proposed SADL.Although sparse representation based dictionary learning methods have obtained remarkable results in face recognition,these methods all act on image descriptor level.Dictionary learning is not well exploited in building of image descriptor.Therefore,a bottom-up dictionary learning based classification for face recognition(BUDLC)method is proposed,where a patch dictionary and a classification dictionary are well learned in the building of image descriptor and classifier,respectively.The two dic-tionary learning stages are connected through space pyramid matching.Good perfor-mance has been achieved in face recognition with and without occlusion.