Discriminative Dictionary Learning Algorithm In Sparse Coding Based On Classification

Sparse coding is a method that uses over-complete dictionary atoms to linearly represent image. For each code coefficient, the non-zero elements account for only small part of all elements, which represent the sparsity. Sparse coding algorithm based on classification in dictionary learning optimization process achieves an optimal weighting of the dictionary atoms for each feature, which improves the quality of the dictionary and accuracy of image classification. However, sparse coding in image classification is imperfect. How to effectively improve the accuracy of the image classification in sparse coding is the direction for researches.The paper mainly study the discriminative dictionary learning algorithm which is based on the worse quality of dictionary and classifier. By improving the relevance between dictionary atoms and increasing the accuracy of classifier, the algorithm makes the similar features produce similar sparse coding and raise the classification result. The main contents of this paper are as follows:Firstly, a discriminative dictionary learning method is proposed based on M nearest neighbor. In the label consistent K-SVD algorithm, the trained classifier is a simple linear classifier, which can’t cluster the same or similar dictionary atoms accurately. M nearest neighbor method can not only reduce the size of the dictionary, but also decompose the classifier into various small classifiers with high accuracy. The new algorithm clusters similar and same dictionary atoms more accurately, so as to achieve better classification result.Secondly, a discriminative low rank dictionary learning method is put forward. Through the low rank representation, we can update the dictionary to obtain a more pure dictionary, which enhances the robustness of the dictionary. The paper proposes the M nearest neighbor’s discriminative low rank dictionary learning(MLR) algorithm, which obtains better sparse solution and pure classifier with higher accuracy to improve the effect on clustering the dictionary atoms.The above two algorithms are tested on the Extended Yale B、AR and Caltech101 data sets, the result of which are compared with those of the state-of-the-art methods including submodular dictionary learning algorithm, Fisher discrimination dictionary learning algorithm and the label consistent K-means singular value decomposition algorithm. The experimental results show that the proposed algorithm outperforms the other algorithms.