Image Retrieval Research On The Uniform Region And Weber Local Descriptor

For the increasing digital image data, the image information available is only usefulif one can access it efficiently, so there is a growing need for effective image retrieval,which can effectively store and manage the image database and quickly retrieve a desiredimage according to the user’s needs. The image comprises text and pixel information,text information uses keywords to depict the visual content of image, and pixelinformation reflects the objective visual content the image contained. Text based imageretrieval(TBIR) need to be marked by keywords, which expenses large and subjectiveambiguity, and automatic annotation based on machine learning exists its own limitations,whereas content based image retrieval (CBIR) can overcome the problem. So CBIR isone of the hot shots in the information retrieval field.The thesis first discusses on two parts: one is basic knowledge about CBIR, the otheris weber local descriptor(WLD) and the relevant local descriptor weber local binarypattern(WLBP). Then we do research on image retrieval approach in the direction of thelow-level feature extraction and region based image retrieval, combining it with WLDlocal feature. The research work can be summarized as follows:1) For the existing problems of the traditional block color histogram and combined withhuman visual characteristics, two image retrieval methods based on the uniformregion are put out, one is WLD image block classification weighted image retrievalmethod, which concerned about different human visual attention on different blocktypes; the other one is uniform elliptical ring partition method, which pays moreattention on the visual center and has different attention degree on different regions.Both WLD image block classification weighted image retrieval method and uniformelliptical ring partition method are based on the integrated color and texture feature,and we integrate color and texture feature without extra internal normalization orexternal normalization. 2) The dimensions of WLD and WLBP local descriptor are high, which limits theirapplication in image retrieval. The improved WLD/LBP descriptor can use lessfeature dimension but ensure the retrieval efficiency. Last we test on two imagedatabases using different similarity metrics, and all the results show that theimproved WLD/LBP descriptor is better than the standard WLD, LBP, WLBP.