Research On Content-Based Image Retrieval

With the rapid development of the multimedia network technology, theapplication of the image becomes more and more extensive and Content-basedImage Retrieval (CBIR) becomes one of most active researches in multimediaretrieval field. In order to analyze the information of an image, the CBIR systemalways analyzes the color, texture, shape, and other low-level image features,taking feature vectors as retrieval index. Up to now, the main CBIR method issimilarity matching based on multi-dimension feature vector of image. Extractingfeatures from image and similarity match are the key issues in CBIR.In this paper, we extensively studied the national and international materialson CBIR systems, discuss the research status and trend in content based imageretrieval problems, and analyzed and implemented some methods of extractingimage content in details. In this paper, how to extract the feature of shape andspace-interrelation and how to integrate the feature of color, texture, shape andspace will be discussed detail. This dissertation discusses the visual content basedmethod based on image feature extraction and similarity matching. We designs animage retrieval system based on image content, the experiments show that theshape features and the space-interrelation features are effective in describingimage content, and the combined feature by the four feature is superior to anysingle feature of the color, texture, shape and space-interrelation on retrieval.As important visual information of image, the color feature is often used todescribe image content and has been used broadly. In this dissertation we havediscussed key problems on how to make use of the characteristic of color, whichincludes describing color, obtaining the characteristic of color. The method ofextracting color feature based on image content is analyzed, color histogram builtfrom cumulative distributions of content colors is a main color feature used inimage retrieval. Based on extracting color feature from BMP static image, colorhistogram is adopted. To show color characteristic of image, the method of theHSV color space, which is suitable to the visual characteristic of human, is utilized.Taking advantage of human's feeling on color, it quantifies color sector withunequal interval, and gets characteristic vector, regarded as color feature of image. Texture, as an important component of the human vision, can reflect the depthand surface information of image, and supply the human vision with therecognition and understanding information, is a feature hard to describe. In thispaper, co-occurrence matrix is adopted by extracting texture feature. Shape is another important feature of object in image, has good stability. Inthe retrieval mainly based on similar shape feature, the shape feature shows thefine capability, which the color and the texture feature can't compare with. So theresearch on how to extract shape feature from image is very important. This paperhas analyzed the shape feature and presented a moment invariant based onthreshold optimization method. In this method, we transformed the color image togray image first, and used the moment invariant based on threshold optimization tosegment the gray image, and then extracted the edge of gray image. So far as wehave obtained the edge of gray image, and calculates the seven-moment invariantof edge of gray image. The seven-moment invariant was the shape feature of theimage. In order to make the seven-moment invariant supply the same measure, weshould be make every the seven-moment invariant dimensional uniformity throughGauss-uniformity. The result of experiment shows that moment-invariant possessthe image object's invariant to translation, scaling and rotation, it can describe theshape of image and the space information well, and especially for the image withclear edges and object the retrieval result is good. Spatial relationship is the important feature of object in image and itdescribes the inter-relationship between every object in image. When the shapeand size of object is far less than the distance of the two objects, the retrieval byspatial relationship is more effective. In this paper, a new content-based spatialrelationship retrieval method is proposed. In this method, we translate the RGBspace of color image into HSV space first, and then quantifies color sector withunequal interval, and get feature vector, take every bin of H as a state, then wecalled the gray image of H as state matrix. Using the Z-scans, the state matrix istransformed into 1-dimensional state sequence. Here we suppose that the...