Image Classification And Semantic Indexing Based On Fuzzy Support Vector Machine

With the increasing amount of multimedia data, content-based image retrieval attracts many researchers of various fields. In which, many standing achievements have been made, and on which the research is developed rapidly. Image semantic indexing also becomes an important and challenge task of the field of content-based image retrieval. Since digital images and videos are becoming a major source of multimedia data, image semantic indexing is a very imperious demand. The traditional manually image indexing is not only infeasible when the number of images is fast increasing, but also is easily influenced subjectively by the operator. Therefore, image semantic indexing automatically is important for image semantic retrieval.Support vector machine (SVM) is a basically two-class classifier and especially performs well when there is no overlap between classes. For the n-class problem in image classification, SVM convert it to n two-class problems, which is also called one versus rest (1-v-r) problem and one versus one (1-v-1) problem. Unclassifiable regions exist in both the extended method. So we introduce fuzzy support vector machine (FSVM) to solve the unclassifiable regions by defining a member function.Due to FSVM's performance in multi-class classification, it's used for image semantic indexing based on image classification in this paper. We propose a new method for image semantic indexing. In order to understand the image better, use the human understanding of images for reference, which means the first level is focused on interest, typical, semantic region in image, then, the layout among these regions. We introduce a 3-level image pyramid structure for image semantic indexing.In this paper, FSVM is employed for training the concept model in the concept model library and calculate the likelihood between an image and a concept. We index the images with the concepts according the likelihood which is adding up according to the weighted image pyramid and realize a kind of intelligent, coincident in visual perceiving image understanding and indexing.