Detecting prominent objects for image retrieval

This paper presents a novel method for detecting objects in natural images in a non-restricted domain. Our purpose here is to capture meaningful information in terms of user semantics on the visual level for efficient image retrieval. The major obstacle in developing such methods is the difficulty of accurately segmenting the image into its prominent regions. To overcome this difficulty, we developed a vector-valued inhomogeneous diffusion model that uses multiple features. In this model, we define the gradient threshold and thus the conductance parameter as a function of the texture and/or color gradient varying by evolving diffusion. The method is evaluated for 20,000 natural images taken by both professionals and amateurs and it is proved to be very robust and effective for obtaining sufficient region-based image description and thus facilitating intelligent and user-friendly image retrieval applications.