A novel fusion approach to content-based image retrieval

This paper proposes a novel fusion approach to content-based image retrieval. In our retrieval system, an image is represented by a set of color-clustering-based segmented regions and global/semi-global edge histogram descriptors (EHDs). As a result, the resemblance of two images is measured by an overall similarity fusing both region-based and global/semi-global-based image level similarities. In our approach, each segmented region corresponds to an object or parts of an object and is represented by two sets of fuzzified color and texture features. A fuzzy region matching scheme, which allows one region to match several regions, is then incorporated to address the issues associated with the color/texture inaccuracies and segmentation uncertainties. The matched regions, together with the simple semantics for determining the relative importance of each region, are further used to calculate the region-based image level similarity. The global/semi-global EHDs are also incorporated into our retrieval system since they do not depend on the segmentation results. These EHDs not only decrease the impact of inaccurate segmentation and but also reduce the possible retrieval accuracy degradation after applying the fuzzy approach to the accurate segmentation for images with distinctive and relevant scenes. The Manhattan distance is used to measure the global/semi-global image level similarity. Finally, the overall similarity is computed as a weighted combination of regional and global/semi-global image level similarity measures incorporating all features. Our proposed retrieval approach demonstrates a promising performance for an image database of 5000 general-purpose images from COREL, as compared with some current peer systems in the literature.