关键词: 梯度, 轮廓, 纹理, 视觉注意, 显著性

Abstract: Accurate extraction of contour from cluttered natural scenes still remains to be solved, while traditional segmentation based on gradient map is facing a bottleneck. This paper analyses the visual diversity of contour and texture in gradient map, and illustrates the necessity of saliency detection. It introduces visual attention and utilizes residual spectrum to acquire gradient saliency map, which enhances the gradient response of contour and inhibits gradient response of texture. It presents the implementation of the method. Compared to other three algorithms, it demonstrates that edge segmentation and detection based on gradient saliency map inhibit texture edge and improve the performance of contour extraction significantly. By varying the parameters, it shows that the algorithm obtains a certain extent of robustness to parameter variation.

Key words: gradient, contour, texture, visual attention, saliency