基于梯度显著性的轮廓提取方法

Contour extraction based on gradient saliency

在复杂的自然场景中将轮廓准确地提取出来一直是一个难题,传统的基于梯度图像分割的方法在性能提高上遇到瓶颈。分析了梯度图像中轮廓与纹理的视觉特性,说明了显著性检测的必要性;引入视觉注意机制,利用残余谱得到梯度显著图,突出了轮廓的梯度响应的同时抑制了纹理的梯度响应,证明了显著性检测的可行性;详细介绍了算法实现流程。通过与其他三种算法进行对比,证明基于梯度显著图进行边缘分割和跟踪,有效地抑制了纹理边缘,轮廓提取性能得到明显提高;通过调整参数设置,验证该算法对参数变化具有一定程度的鲁棒性。

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 textture. 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.