Abstract
In a content-based image database system, forground objects are usually the ones to be retrieved. Therefore, it is desirable to extract foreground objects from images and then store their features for query comparison. This thesis puts emphasis on automatically separating foreguound objects form background such that object retrieval can be effectively achieved. The task is easy for human beings because they can easily recognize objects. However for computers, object recognition is still rather difficult. In this study, we focus on images that are taken from cameras. The foreground objects usually have sharper edges and clearer textures. We first segment an image into regions by a hierarchical coarse-to-fine segmentation using the edge magnitudes. Then we determine for each region whether it belongs to foreground by the information along the region boundary and within the region. Some experimental results are prepared in the thesis.