Perceptual segmentation algorithm and its application to image coding

The success of the recursive shortest spanning tree (RSST) algorithm for segmentation is essentially due to the hierarchical nature of the data structure that it imposes. The homogeneity of regions is used to generate the shortest spanning tree, which subdivides an image into regions hierarchically. However, the subtrees generated by the RSST algorithm do not properly represent the regions in the order of perceptual significance. Less significant regions may be generated before significant ones are defined. This contribution remedies this aspect by creating a novel combination of the homogeneity information from within regions and the saliency information from their contours. The saliency of a region´s contour is a measure that quantifies the perceptual importance of the corners on the contour. It is determined by the intensity gradient and the geometric characters of the contour pixels. By combining the homogeneity from the original image and the saliency from the saliency map, this method is able to subdivide an image into regions in an order closer to the sequence determined by their perceptual significance. Its application to image coding at very low bit rates results in better subjective image quality.