The Research And Implementation Of Technology Of Semantics-Based Image Clsaaification

With popularizing of the digital products, more and more researchers paid attention to image classification technology. Image classification technology has huge potential in research and market. Those big Internet search companies, like Google and Baidu, all were committed to the research of image classification. Existing image classification and retrieval system was based on the similarity of the image visual features. But as the users, people classify images by the semantics of the images. The purpose of this paper is how to make the computer get the semantics of the images like humans.The main content of this paper included three parts.1.) According to SURF feature descriptor doesn’t contain color information, it published a novel SURF feature descriptor based HSV color quantization matrix. Experiments showed that the new descriptor can describe the image more comprehensively.2) This part is about visual vocabulary. The visual dictionary is the bridge between image high semantics features and image low-level content features. The key step of generating the visual dictionary is clustering the visual feature vectors. For the shortcoming of the K means clustering, it published a K means clustering algorithm based on PSO optimization for generating visual dictionary. Experiments showed that the new visual dictionary generation algorithm has more global optimum characteristics than the old algorithm.3) Based on the study of part1&2, this part studied the high semantics feature. It focused on the probabilistic generate models-pLSA&MMI which used mutual information concept. The simulation of these models shows proof that MMI is better than pLSA in accuracy and convergence rate.In the last, it set up an image classification system based on those algorithm we studied.