Research On Relevance Feedback In Content-Based Image Retrieval

Content-based Image Retrieval (CBIR) technology uses of image content features (color, texture, shape, spatial relations and etc.) which are analyzed and extracted automatically by computer to achieve the effective retrieval. In the meantime, CBIR conquered the defects of traditional text-based image retrieval technology, such as strong subjectivity and heavy workload. However, this technology still faces much difficulty to bridge the semantic gap between image semantic features and the lower features, resulting in the fact that the extracted content features still mainly centered upon the lower features such as color, texture and shape. Therefore, it is still an unsolved problem about how to integrate with semantic features to achieve better connection between the lower physical features and the image content for efficient retrieval.In this dissertation, we analyze the Content-based Image Retrieval technology and several algorithms of the relevance feedback in CBIR. Our emphasis is the Feature Re-Weighting approach for relevance feedback in CBIR. After simulating the two classic algorithms proposed by Rui and Aksoy, we propose a novel Feature Re-Weighting approach for relevance feedback. This approach make use of the un-labeled images and the negative images to enlarge the number of the training samples, introducing the concepts of the positive error and the negative error to judge all the positive and negative images which are qualified by adding some limits. This scheme solves the problem resulting from the shorting of the training samples to some extent and simultaneously considers the asymmetry between positive and negative samples reasonably well. Experiments show the noticeable precision, stability improvement from Aksoy.On the basis of studying Feature Re-Weighting approach for relevance feedback in CBIR, we propose the algorithm of combining the lower-feature and semantic to retrieve images. In this dissertation, we use a feature matrix and a semantic relevance matrix which is established by long-term learning the log of the feedback offered by users, then optimize the semantic relevance matrix, and finally, combine the lower-feature matrix and semantic relevance matrix to retrieve images. This approach achieves the estimation of the similarity between...