Research On Image Classification Based On Sparse Repr Esentation And Deep Dictionary Learning

In recent years,with the development of traditional sparse representation face recognition methods,research and application of sparse discriminative dictionary learning algorithms are also increasing.Due to its good learning effect,many excellent results have been achieved in image classification tasks.These single-layer dictionary learning methods can learn strong discriminative dictionaries for specific class of samples,but for the case where the data dimension is too high or the number of samples is large,the information contained in this single-layer dictionary is far from the classification task not enough.In order to obtain more effective image information of complex image sets,in view of the advantages of deep learning methods in the field of image classification,Snigdha et al.Proposed a deep dictionary learning method to define the sparse representation coefficients in traditional single-layer dictionary learning methods as features.The original sample obtains the dictionary and features of the first layer,then obtains the dictionary and features of the next layer based on the features of the first layer,and learns in turn to finally obtain the deepest dictionary and features.This method not only learns a deeper dictionary but also obtains more diverse and representative deep features.This deep structure feature is more beneficial to the classification task of complex images,and show a better classification effect than single-layer features.Although the deep dictionary learning method can obtain deeper-structured dictionary and features,considering that the increase of the number of dictionary learning model layers will cause the original sample information to be easily lost,but it is not conducive to the improvement of algorithm classification performance.Therefore,based on the theories of sparse representation and deep dictionary learning,this thesis combines the graph structure information and category structure information of the original sample,so that the deep structure features obtained through the deep dictionary learning method not only have sparse characteristics,but also have discriminativeness for complex image classification tasks,and merge L_2,1-norm to increase the ability of the algorithm to adapt to the rotation changes of the characteristic object.Then verify the effectiveness and superiority of the improved algorithm on the public data set.The main contents of this article are summarized as follows:1)A deep dictionary learning algorithm based on graph structure is proposed.Use the original sample information and its local adjacency relationship information to construct the graph structure,and apply the graph regularization constraint to the deepest sparse features,so that the obtained deep features and the original sample have a similar adjacency relationship in the geometric structure.Try to prevent the original sample from losing the detailed information contained in it after learning through the multi-layer network,and at the same time improve the algorithm's ability to reconstruct the original sample,so that the obtained deep features have better discriminative ability to enhance the classification effect.And verify the effectiveness and superiority of the algorithm in the classification task on the face database and handwritten digit database.2)A deep dictionary learning algorithm based on label consistency is proposed.Based on the category structure information of the original sample,construct discriminative sparse coding,and apply label consistency constraints to the deepest sparse features,so that the acquired deep features have a similar category structure to the original sample,and maintain consistency between the classes.After deep learning,the category information useful for classification is lost,and deep features with class discrimination are obtained.And verify the effectiveness and superiority of the algorithm in the classification task on the face database and handwritten digit database.3)An algorithm that combines L_2,1-norm with the above deep dictionary learning model is proposed.Due to the rotation invariance of the L_2,1-norm,the use of the L_2,1-norm toimpose row sparse constraints on deepest features can not only better mine the key features of the original sample,but also further make the algorithm have a certain ability to cope with the rotation changes of the feature object.Finally,verify the advantages of joining theL_2,1-norm.