Image Semantic Annotation Method And Its System

With the development of the technology of network, multimedia, database and the popularization of Internet, People's strong demand of multimedia data such as images. Traditional retrieval technology like image retrieval based on text cannot satisfy people's demand completely. Though the retrieval based on content solves the feature of people's vision of image, but it cannot solve high level semantic image retrieval completely. Then it is urgent to establish semantic description and retrieval mechanism of image. "A study on image semantic annotation and its system implementation" is selected as the theme of this paper.CSIR is an image retrieval system which is be able to provide users with content-based and semantic-based image retrieval. Based on this system, this paper analyzes and studies a real-time image semantic annotation methods, and design and implement the core of the semantic-based image retrieval system--Image Semantic Annotation System. The main contributions of this thesis are as follows:(1) Analysis and research the method RALIP (Real-time Automatic Linguistic Indexing of Pictures). Its main contents are: non-European space image similarity measure, machine learning algorithm D2 clustering, based on the probability of mixture model, as well as model-based real-time annotation method.(2) Propose a fusion method of annotation results using uncertainty reasoning theory. This paper does a simple modification for RALIP, it first traines a model based on single feature, then fuses the annotation results based on different characteristics. Experiments show that the method usually can annotate an image good.(3) Design and implement a real-time image semantic annotation system. The system is implementation of modified method RALIP. It is composed of two parts: model training subsystem and image annotation subsystem. They are relatively independent, the former can be off-line and train images to obtain model; the latter can be online and annotate an image.