Research And Implementation, Based On Relevance Feedback And The Characteristics Of The Image Content Retrieval System

With the huge increase of information in twenty-one century, effective search for information is becoming a hot area. But traditional keywords-based image retrieval no longer satisfies with people's desire in image retrieval area, Content-Based Image Retrieval (CBIR) emerges in this situation. CBIR is a system which searches most similar image in image database and displays them to user through comparing images'vision features like color feature or shape feature, which extracted by CBIR system from the images. CBIR system uses image's feature to represent one image, and the input image is an obvious image example, the outputs are images which is most similar with the input image. All of the searched images are ordered by degree of similarity, the more similar with the input image, the more precedent in the output images. Then users can choose the image they think similar with the input image. So, CBIR easily completes presenting image feature, extracting this feature, recognizing this feature which user hardly complete. CBIR system is used in a small area such as library digital searching system; it has a long way to popularization.After introducing the development of CBIR, this thesis presents the foundation of digital image processing and Otsu arithmetic which is used to choose a threshold to divide one image according gray-level.Next, this thesis introduces how to extract color, shape, texture and space feature, calculate similarity degree, process inside normalized and outside normalized, and evaluate the search result. Then proposing a search method based on combining color, shape and space feature to search image. This thesis uses HIS color space to extract color feature, moment invariable to extract shape feature, and divides one image to nine rectangular parts, and then put different weight to different parts. It calculates combined features by normalizing the color feature and shape feature and global similarity by adding different feature with different weightFor resolving the semantic gap in CBIR system, this thesis introduces relevance feedback, and proposes a method that combines database and cluster method to relevance feedback based on Rocchio's relevance feedback, makes relative feedback has remembering function.For reducing response time, this thesis clusters color feature by K-mean cluster method and the best cluster number is decided by DB Index rule automatically, and it uses cluster's center as index to search database. This thesis's experiment proves it reduces search time to a large extent.Finally, this thesis implements a prototype system based on methods it proposes and makes contrast experiment based on these methods. The result indicates that methods this thesis proposes have feasibility and good practicability.