Research And Implementation Of Combining Visual And Semantic Image Retrieval Technique

Image retrieval is a hotpot in digital library. Two effective ways has been proposed to research: one is visual-based image retrieval technique and the other is semantic image retrieval technique. The first way would be matched by visual content of image, such as color, texture shape etc, but it lacks perception of people. The second is to get image semantic information via different ways, then locate image. But several words can't accuracy description an image. The two ways would be combined.Owing to some unsolved technological problems, few content-based image retrieval systems are put into use, especially lack in the field of object-oriented image retrieval. This paper is an explosive discussion on CBIR in accordance with military digital library. We combined methods of semantic-based and visual-based retrieval. The paper analyzes key technologies of image retrieval in low-level feature and semantic-level, i.e. visual feature extraction, multi-dimensional indexing. similarity metric equation, semantic description and algorithm, then color-based and space retrieval algorithm and main object-based color clustering index method are proposed. These can improve accuracy and reduce user's time. Relevance feedback is also applied in the paper. On one hand, it enables the system catch users' query intention on line by adjusting its similarity criterion automatically; on the other hand, it update weights between key words and images and fill the semantic networks, thus the system can "understand" the user's mean and study in long-term. In this paper, a new image retrieval algorithm in digital library is proposed, it based on the combination of visual features and semantic contents, in addition to the relevance feedback technique, give a dynamic similarity metric equation. Users can provide their query requirement in different manners, the best retrieval result will be presented for the low-level feature and semantic network cooperating properly. Experiments on our prototyping system show that it gets higher retrieval rate than only by visual features.