A K-MEANS And LDA Based Discriminative And Compact Dictionary Learning Method

In recent decades, sparse coding has become one of the most important concepts in the field of the signal processing. Sparse coding is an unsupervised learning algorithm, which aims to look for a set of basis vectors to represent the data efficiently. The over-complete dictionary can help to find the hidden structure and relationship within the original data.Dictionary learning is an approach to learn attributes from a set of training samples and K-SVD is a common algorithm to learn an over-complete dictionary. The over-complete dictionary contains the powerful data representation capacity, which is widely used in machine learning, neuroscience, signal processing, and statistics.However, the over-complete dictionary also brings some problems. The large size of the dictionary leads to the low computation efficiency and some of the items have no effect or even bad effect for the recognition. For this reason, more and more compact dictionary learning method are proposed in recent years and perform very well in some datasets. Based on the above, a LDA(Linear Discriminant Analysis) based compact dictionary learning method is proposed. Meanwhile, a K-MEANS based improved algorithm is also proposed to solve some problems of the LDA method. The main contributions of this paper can be highlighted as follows:(a) A LDA based compact dictionary learning method is proposed in this paper. The sparse representation obtained from the over-complete dictionary usually affects the recognition results directly as it represents the original data and usually serves as the input of the classifier. LDA models the dimensionality reduction as a mapping problem aiming to minimize the intra-class scatter and also maximize the inter-class scatter. The mapping matrix usually can effectively map the sample data to be more discriminative. We apply the LDA-based mapping method to transform the original coefficients to be more discriminative for compact dictionary learning, resulting in high intra-class similarity and high inter-class dissimilarity of the compact dictionary coefficients for better classification.(b) A K-MEANS based improved algorithm is proposed to solved the over-fit problem of the LDA based method. The LDA based method discussed above constraints that the size of the compact dictionary is smaller than the class number. In order to improve the performance of the LDA based algorithm, an K-MEANS based improved algorithm is also proposed. The improved method releases the restriction and achieves the same or even better performance.The proposed method can learn a small size of compact and discriminative dictionary with global optimization. By comparing the recognition accuracy, the compactness and the purity, the proposed approach achieves a state of the art performance on various public action recognition and object classification datasets. Experimental results demonstrate that our approach also outperforms several recently proposed compact dictionary learning methods on object classification and human action recognition.