Complete contours formation via alignment interpretation of illusory edges

This paper presents a statistical approach to the formation of a complete set of “meaningful contours” conformal to human visual perception, characterized by preserving both illusory edge pixels and their corresponding physical edge pixels through alignment interpretation. The method mainly consists of two stages: (i) learning alignment significance (LAS) based on the evaluation of orientations neighboring a physical edge pixel, specifically, alignment of a single orientation is assessed with an elliptical sampling mask, plural mask shapes are determined according to the entropy contained in the mask, followed by borrowing the concept of Maximum-Likelihood Estimation (MLE) to pick out the optimal one representing the alignment significance. (ii) Depicting complete contours by subjecting the result of (i) to a voting scheme to produce proper salient measures for each of physical and illusory edge pixels. Experimental results are provided to justify the effectiveness of depicting human perceived contours either on synthesized and real images. Test results on a dataset of 40 natural complex images are compared to show the fairly superiority of our method others.