Study On The Approach Of Fusion Relevant To The Feedback Technology And Feature In Content-based Image Retrieval

At present,Content-Based Image Retrieval(CBIR) is becoming a hot research topic.It is a retrieval technology based on the vision features,such as the color,texture and shape of the image. It is different from traditional retrieval technology based on the text.This technology directly analyzes image content and extracts image features. These contental features to be built index and be used in retrieval. But, computers have limited ability in understanding and analysis images and cannot precisely understand semantic information of images. This thesis is mainly focused on bridging the gap and on research of relevance feedback and feature fusion theory and algorithms for image retrieval.The thesis, first, researches the low-layer physical feature of image. Along with the analysis of HSV color space, we find out that color at the neighbourhood of boundary of color quantization in H channel is gradual changed and intermixed. The thesis presents a novel color quantization method—fuzzy color quantization based on H channel. Adding fuzzy information in H channel makes color quantization more reasonable for human's vision model. This method solves the so-called boundary problem in color quantization. The extensive experiments show that the fuzzy color quantization based on H channel algorithm is effective and robust.The method of multi-features cooperation retrieval has been recognized as an effective method in improving image retrieval performance. From another viewpoint, CBIR can be viewed as sort problem. The second part of this thesis is carried out to deal this problem. The thesis presents a novel Borda count method that integrates correlative information between multi-classifiers. The algorithm considers that in image retrieval, multi-classifiers have intensively correlative information, and punishes or encourages images that were sorted by respective classifier according to a proposed strategy, then, and re-sorts the images. This algorithm realizes fusing effectively results of multi-classifiers by preserving the result of "good" performance of classifiers and oppressing the result of "bad " performance of classifiers. Contrastive experiment shows that the algorithm is effective.Relevance feedback is an interactive technique in the process of CBIR. Support vector machine(SVM) is developed based on the statistical learning theory in past few years. Traditional SVM as machine learning in CBIR only considers relevance feedback as pattern class problem. While, in massive image data, images reflected into feature space are not linear separable. Simply pattern class method which be used as relevance feedback cannot achieve ideal result in image retrieval. The last part of this thesis presents a novel twice constrained distance algorithm(TCDA). After getting two class images classified by SVM, TCDA re-sorts classified images by the distance between the clustering center point of relevant images and all images in database, then presents the result to users. Extensive experiments show that the novel algorithm is effective and has ideal learning result than before algorithm.