Nowadays, with the fast development of information industry, human beings have been in the information age, in which multimedia and network techniques are in the state of exponential growth. In order to effectively store the mass data, people inevitably need to build massive databases. Therefore, how to search the images people need from massive image database more accurately and more quickly has recently become the common concerned problem of researchers in multimedia technology at home and abroad. Just against this background, content-based image retrieval techniques (CBIR) has attracted extensive attention and become a hot research field.Feature extraction is the core content of CBIR. The merits of feature extraction algorithm directly affect the retrieval performance. By intensively studying and analyzing traditional feature extraction algorithm, this paper uses edge detection operators to extract the edges of image and classical color multi threshold method to segment image and extract the image shape. On this basis, this paper makes some modifications to the existing color histogram, Hu moments and Curvelet transform algorithm. When modifying Curvelet transform algorithm, man value, variance and skewness three parameters together indicate the energy distribution in each multi scale Curvelet decomposition level. Compared with the original algorithm, the modified algorithm can describe more detailed image texture. Result of testing10types of images in database shows that the average accuracy of modified algorithm is increased.In order to remedy the limitation of using single feature to description image, this paper also further studies the image retrieval based on feature fusion. To better fuse color, shape and texture three features thereby to improve retrieval performance, this paper combines analytic hierarchy process (AHP) and relevance feedback technology. Based on the statistical data obtained from the modified single feature retrieval algorithm, AHP is used to set up the initial weight of single feature. After the first retrieval, relevance feedback mechanism is built. Users can evaluate the retrieval result and according to the evaluation, the system can automatically modify the weight to match users’retrieval intention. Thus human-computer interaction is realized. The experiment of retrieval of image retrieval system based on feature fusion done in MATLAB R2010b, shows that combing using AHP to set up the initial weight of three single features with using relevance feedback to automatically adjust weight can not only improve the retrieval performance of image retrieval system, but also can effectively reduce the feedback times and improve the efficiency of retrieval system. |