Sparse Concept Coding Based On Information Entropy And Its Application In Image Classification

Image classification as one of the important research direction in computer vision. The development of this technology has the vital significance to the real working life and the development of the society. Manual for text annotations is proposed in the early 1970’s and sparse representation is put forward in the 1990’s. Based on the concept of sparse coding two new methods to improve this technology are proposed. Specific work is as follows:1. A sparse concept coding method based on information entropy (ESC) is proposed for image classification, which is embedded in the information of the information entropy of the data. In the process of mapping data, this method overcomes the disadvantages of artificial adjusting the scale parameter σ, using the iterative adaptive estimation of the optimal scale parameter σ. Then by formula derivation that scale parameter σ and information entropy H (x) the relationship between the characteristics, scale parameter optimization using this feature. According to the optimal scale parameter sigma obtained similarity matrix to solve the feature vector space. Finally, through the vector based learning and sparse said learning representation and classification of the sample. The effectiveness of this method in image classification is verified by comparing with other methods.2. A sparse concept encoding based on information entropy weighted (WESC) is proposed. Combined the separability criterion with information entropy, namely in the class separability criterion based on considering the probability distribution of the sample, that is the separability criterion lack the probability distribution information of make up as well as to deal with the weighted information entropy of the sample. Therefore, it can improve classification accuracy and Normalized Mutual Information.