Research On Image Retrieval Based On The Curvelet Transform

With the development of computer science, internet and multi-media technology,many great changes for human sociality, including economy, daily life, work, study etc.,have taken place. With rapid improvement in the ability of image collecting, explosivegrowth in the amount of image storage, how to retrieve the useful image informationquickly and exactly from vast images has become a research focus.Content-based image retrieval(CBIR) can make use of the low level visional featuressuch as color, texture, shape and edge etc. It has been widely researched because of itsadvantages on overcoming many shortcomings of traditional retrieval appraoches. InCBIR, multi-resolution analysis can provide local time-frequency expression, and thepossibility distribution models can fit the distribution of subbands in frequency fields, soit is the trends that make use of the combination of possibility distribution models andmulti-resolution analysis. As a good multi-resolution analysis, wavelet was often used intexture image retrieval, but wavelet was found lack of directionality, curvelet wasresearched as an improvement because of the better directionality and more statisticalinformation than wavelet. So far the features used on curvelet are low level statisticalfeatures which are simple, and the image texture information extracted are poor. Forimproving the retrieval ability we introduce the high level statistical features to CBIR.In this paper, we computed the generalized gaussian density and gray level-gradientco-occurrence matrix on the sub-bands of curvelet. Extensive experiments demonstratedthe effectiveness of these two methods. The main works are as follows:1. Based on the multi-directionality of curvelet, we executed generalized gaussiandensity model on its sub-bands, and used the Newton-Raphson iteration to compute theparameters of probability density functions, then taken the Kullback-Leibler distance asthe similarity measures on the parameters, we obtained some better image textureinformation, the comparison of retrieval results produced by several algorithms wascarried out on texture image library. After that, a scheme by weighting on the texture andcolor was proposed and tested on color images, good retrieval results are obtained.2. Gray level-gradient co-occurrence matrix(GLGCM) was applied on curveletsubbands, we got a set of effective image texture features by computing gray mean, graydeviation, low gradient dominant, energy, relativity and opposite differential moments onGLGCM. Finally, the same weighting scheme was applied on GLGCM texture featuresand color features, and good retrieval results were obtained.