An integrated segmentation technique for interactive image retrieval

Many content-based image retrieval (CBIR) systems utilize image segmentation for enabling the user to perform object-level database querying. We propose an integrated segmentation technique for interactive image retrieval, that is reasonably accurate and fast. An initial over-segmentation is generated by finding the dominant color modes in the global histogram of the image using the mean-shift algorithm. Edge-based processing is performed at the initial segment boundaries to merge non-obvious segments. Finally segment shapes are regularized using a Hopfield (1985) type neural network to improve their perceptual quality. A scalable implementation is presented for ensuring fast serial execution of the Hopfield network. The entire segmentation process takes less than 10 seconds to segment 128/spl times/192 stock photos on a standard workstation.