An object-based approach to similar image retrieval

This paper proposes a new object-based approach to similar image retrieval. First the salient regions of an image are detected by using a novel segmentation method based on a multi-scale inhomogeneous diffusion model applied to color and texture features. Each detected region is then represented by a feature vector composed from the characteristic color, texture and shape features of a region whose features are invariant under rotation. Scale invariance is also addressed for the color and shape properties. An R* tree-based indexing scheme is applied over the feature space to ensure efficient searching. By applying a suitable user interface, the method can handle sub-image queries and object-based queries with regard to a certain object or objects in the input image specified by the user. Experiments conducted on a large number of images taken from photo-CD data and collected from the Internet, show that the method performs well for a large variety of natural images.