Semantic-based Image Annotation System Design And Realization

Today text-based image retrieval technology is very mature, content-based image retrieval is also perfect, however, there is a big gap between the image content and the semantic of the image represented (Semantic Gap). Text can be more directly express the people's search intention than the image content, but in the face of such a large image scale, manual annotating is not sufficiently precise and time-consuming. Therefore, let the computer understand the image and then auto-annotate has become an urgent task. Due to people's understanding of things is a multi-level, from the abstract to the concrete, therefore, to meet the needs of users'inquiry at different levels, it is necessary to achieve the multi-level image annotation. In this article, we first introduced salient object as the elements of describe the middle image semantic, the salient object a set of image regions which can represent the physical objects in real life. Second, we have introduced the concept ontology to describe the image semantic framework, we organize the hierarchical structure of image concepts in accordance with the structure of understanding the nature of things in ontological. This work is conducive to the concept reasoning and the latter image retrieval. For the creation of the concept ontology, we have adopted a semi-automatic method: First of all, we use the WorldNet to analysis the context relationship among the tagging vocabulary of LabelMe image library, and then precede it manually. When annotating image, firstly, we segment the image with the image segmentation method based on mean shift, which use the iterating vector to make the points from the same region converging to a fixed point. Secondly, we can achieve the segmentation of color pictures by aggregating and fusing these points into several small regions, i.e. the object of the picture. Thirdly, after processing these objects by pruning marking and so on, extract the features from them. Fourthly, use the support vector machine which has been trained to classify the image regions. Finally, use the Bayesian Networks to achieve hierarchical probabilistic image concept reasoning and finish the multilevel image annotation.