Image Semantic Annotation Based On FSVM

The images as the information carrier are paid attention to by more and more people, due to the large amount of information contained and intuitive form of expression. So image retrieval as a hotspot is valued by domestic and foreign scholars. Image retrieval technology undergone three stages:text-based image retrieval on the 1970's, content-based image retrieval on the 1990's, and semantic-based image retrieval now. Semantic-based image retrieval technology can overcome the defects in the two retrieval technology. The images can be established contact with the semantics underlying characteristics of the image by semantic-based image retrieval technology. When user input the demand, images can be found from the semantic space. A significant difference between semantic-based image retrieval technology and the two technologies is that images are annotated by semantics and images are searched based on semantics. So the important link of semantic-based image retrieval technology is image semantic annotation that affects the accuracy of semantic annotation to retrieve the final results.General approach to semantic annotation is that the images are annotated directly through characteristics of the images. When making the semantics extraction and semantic annotation of images, the semantic gap problems scholars who regards the method solved the problem is as current academic research focus. Scholars generally use the idea of machine learning to solve this problem. A set of images annotated as training set are used to train semantic annotation model by which images can be annotated. In this idea, the Support Vector Machine has its advantage in solving the problems about the small amount of samples. The theory of Support Vector Machines and fuzzy Support Vector Machines are studied and a kind of fuzzy membership function is introduced in the FSVM. In the theory of FSVM, the fuzzy membership function as an important parameter indicates the probability that some samples belonging to a classification. After contrast to the commonly used fuzzy membership function, a new function is proposed which consider the distribution around the samples. Commonly used fuzzy membership functions are generally considered the relative distance between a sample and the class center of the samples, and without considering the distribution around the sample. So there are some limitations when they are used. This compact fuzzy membership function proposed in this paper which considers the sample distribution around the sample, can more accurately reflect its probability belonging to a certain classification. When it is introduced in FSVM, this FSVM which has better accuracy of image annotation has better performance than SVM and usual FSVM that use linear or S-type fuzzy membership function. This can be showed from the results.