Some Technologies Research On Content-based Image Retrieval

With the rapid development of multimedia and internet technology, people can get more and moreinformation of all kinds of images. Face to the massive image information, the content-based imageretrieval (CBIR) system which is used to rapidly and effectively search the desired images fromlarge-scale image has become a research hot in multimedia field. In recent years, CBIR is a very hotresearch direction at home and abroad and has been applied to many fields. Up to now，still manyproblems need be solved in this research field. This thesis focuses on CBIR technology and gives somecontributions as follows:Firstly，in order to get rid of the noise impact for image retrieval, two denoising algorithmsincluding a modified bilateral filtering algorithm and an algorithm combining the improved non-localmeans and non-subsampled contourlet transform Wiener filtering are proposed. The former uses wienerfunction to estimate the values of the current and neighbor pixels, based on which the radiometricsimilarity weight values are computed. Thus, noise interference on the weighted coefficient can bereduced. The latter is a denoising algorithm integrating spatial and transform domain. In spatial domainthe highfrequency noises are firstly removed using the improved non-local means method with smallradius of the neighborhood. Then in contourlet domain the de-noised image using non-local meansmethod is re-denoised by Wiener filtering and most of low-frequency noises are removed. In non-localmeans method, three aspects about denoising performance，computational complexity and parameterestimation are improved. The experimental results show that the proposed algorithm can get denoisedimage with higher subjective visual quality and objective evaluation index.Secondly, an image retrieval method based on color-spatial distributing feature is proposed，whichcan effectively overcome the problem that the traditional image retrieval method is prone to lose thespatial information of colors. This mehod not only utilizes the pixel position but also the same colorpixel spatial distribution characteristics. According to visual attention computational mode，theweighted histogram which reflects the pixel position importance is constructed after all pixels areweighed by the pixel color contrast with respect to their multi-scale neighborhoods. In the meantime, the spatial relationship feature of the same or similar colors is considered by colors distributingcohesion. The experiments show that the method mentioned above has high accuracy and its retrievalresults match human visual percept well．Thirdly, in the region-based image retrieval, key problem is to accurately extract region of interest.An image retrieval method based on visual attention model is proposed. Employing visual attentionmodel combining Gaussian multi-scale transform with color complexity measure to create salient map,based on which the salient object is extracted by max-variance method. Simultaneity edges in salientregion are extracted according to the created salient map and the initial edge map by canny detector.Color histogram of salient object and gradient direction histogram of salient edges are fused for therealization of regional image retrieval.Fourthly, an image retrieval scheme is proposed based on color and object semantic concept ofeffective visual area, which can effectively overcome the “semantic gap” between the low levelfeature and high level semantic concept. A visual window including single object is created accordingto global sailent value of pixel and edge information based on distribution of corner points. Then，thehigh level color semantic can be obtained by using the main color of the quantized color image．Andthe high level object semantic is determined by mapping the low level feature using SVM Machinelearning. Finally，image retrieval is implemented by using these two-level high semantic concepts.