Research On High-level Semantic-Based Image Retrieval

The rapid development of computer, multimedia and Internet techniques results in large quantity of image information. Thus how to efficiently and rapidly search out the customer-demanded image from large scale Web database becomes an important problem, which urgently needs to be solved. However a great semantic gap breaks the relation from the visual characteristics of the image to its semantic expression. In order to improve index's efficiency and accuracy as well as diminish the existing great semantic-gap, High-level Semantic-Based Image Retrieval becomes focused area in the image retrieval field.Based on reviewing the current status of content-based image retrieval in these aspects, the thesis does a deep discussion on the image semantic annotation and the relevance feedback technique. It makes analysis and comparison on the existing mainstream semantic annotation model and relevance feedback algorithm from the view of the algorithm, points out the disadvantages of two techniques and their developing directions. The thesis presents a semantic vector algorithm, builds up the network of image semantic keywords, and realizes the composite retrieval of the image low-level feature and semantics characteristics. Through relevance feedback technique, on one hand, it enables catch users'query intention by adjusting its similarity criterion; On the other hand, it passes the information for relevant images, updates weights between keywords and images, and fills the semantic networks. In addition, an efficient image retrieval prototype system is built. Users can provide their query image; the best retrieval result will be presented for the low-level feature and semantic network can cooperate properly. The experiment shows that in few steps of relevance feedback, the recall and precision are improved effectively.Finally, we conclude the whole thesis. The difficulties, hotspots and problem solved in the next step of research are pointed out.