Research On Segmentation And Salient Regions Extraction In Image Retrieval

As the development of multimedia technology, how to organize and search effectively image information has become a valuable research subject with image retrieval as one of the most significant parts. In this dissertation, image segmentation and region of interests extraction techniques are discussed. More specifically, image segmentation is based on support vector machines (SVM) and D-S evidence theory. Region of interests extraction techniques are based on SVM. Three parts are included in the work:1. A novel method of region-based segmentation using SVM and D-S evidence theory is proposed. Image segmentation is a significant preprocessing step for high-level computer vision tasks. As D-S evidence theory is an information fusion theory, its combination with SVM has been a hot area of research. Combination of Support Vector Machine (SVM) and Dempster-Shafer (D-S) theory is applied to the field of region merging to deal with the uncertainty. In the proposed algorithm, SVM is utilized as the identifier, and Basic Belief Assignment (BBA) function is constructed accordingly. Fused BBAs are obtained by applying the D-S evidence theory to the outputs of the identifiers.2. An improved approach of regions of interests (ROI) extraction technique based on SVM is proposed. A new ROI extraction principle is applied in the proposed algorithm, with color contrast and position information included. Compared with the surrounding area, color contrast of ROI appeals to people’s attention. Generally, ROI tend to get together, while the background regions appear in both near and distant areas. The performance of extraction method can be improved by the introduction of color contrast and position information.3. An image retrieval system based on the above is built afterwards. The image information can be described more accurately in the proposed system. To improve the accuracy of similarity measurement, shape context information is incorporated in the system. Moreover, learning scheme is utilized after the retrieval to adjust the importance score of each region, finally narrowing down the semantic gap.