Perceptual saliency weighted segmentation algorithm

The segmentation of an image is critical to the eventual success or failure of image analysis. Most commonly, the similarity or/and discontinuity between pixels in the image is used to group pixels into regions or to produce region contours. The principal difficulty shared by existing segmentation algorithms is that the segmented regions usually do not reflect the objects as perceived by a human being. Unexpected regions are often generated before desired ones are defined. The proposed perceptual segmentation algorithm remedies this aspect based on a novel combination of the homogeneity of regions and the saliency of their contours. Homogeneous regions are defined using the recursive shortest spanning tree algorithm. The saliency of region contours are detected by investigating the intensity gradient and geometric characters of the contour pixels, so as to generate a saliency map of the region contours. By combining the homogeneity from the original image and the saliency from the saliency map, the proposed approach is able to subdivide an image into regions in an order closer to the sequence determined by their perceptual significance.