Color-based Relevance Feedback Image Retrieval Research

With the development of the multimedia and network communication technology, the scale of image database also expands rapidly. The image as a kind of visual and effective form to express information, has already been used in various areas, so it is a concerned problem to search a fast and valid image retrieval.The traditional text-based image retrieval requires a large number of user's comments for the images, for this defect, content-based image retrieval improves the method by extracting image's low level feature on computer, this technique has become a breakthrough in the area of image retrieval, and been experts'research focus home and abroad in recent years. On the basis of existing content-based image retrieval mechanisms, a color feature-color pair which uses the color and shape feature, and an approach which uses color pair as retrieval index are proposed in this dissertation. This approach also utilizes relevance feedback by support vector machine on retrieval results to meet user's requirement. At last, the techniques analyzed in this dissertation for improving retrieval performance are proved in the experiments.The main works of this dissertation are listed as follows:Among the image's color, texture and shape features and so on, the color feature is used the most because of its high stability, so this feature is studied mainly in the dissertation. At first, on the basis of the analysis and extraction of color feature, the relevant theory and algorithm are discussed. Considering kinds of expressions about color feature, a good expression-color pair is found, it is a color feature that can fuse image's color and shape features as retrieval index. During the process of extracting image's feature, some optimization disposals are adopted to raise retrieval accuracy, including using a series of preparation methods to improve picture quality, choosing HSV color space which is fit for human eyes sensory, doing the operation of quantizing and clustering on color feature to reduce feature vector's dimensions. The specific process of color pair retrieval is introduced in detail, and the dissertation has illustrated this method's improvements compared with others, including choosing different image concerns, the invariance for image's changes of rotation, translation and extension, in the end, some experiments are done to prove its superiority.For a further improvement on the accuracy of color pair method's initial return result, the hot technology-relevance feedback mechanism which is to reduce the difference of image understand between user and computer is studied here, and the feasibility of this method is shown. Support vector machine studying a small sample of positive cases is used to search a learning model that guide new retrieval to optimize retrieval system's performance, a experiment result after many times of feedback is given at the end of the dissertation.Based on all the involved theories and means analysis of color pair-based image retrieval which is proposed in the dissertation, a general framework of the image retrieval by integrating technology studied organically is designed , and experiments are carried out repeatedly to analyze its improvements from data.