基于递进扩散模型的显著性检测

Saliency Detection Based on Progressive Diffusion Models

针对现有基于扩散的流形排序算法缺少利用扩散后的结果更新图结构、导致其在复杂数据集上未能充分抑制背景和丢失小尺度显著目标的问题,提出一种基于递进扩散模型的图像显著性检测。首先,采用超像素分割和流形排序相关算法构图,并进行基于伪背景种子节点的扩散,得到第一阶段显著值;其次,将上述显著值采用自适应阈值分割获得前景种子点,再进行基于前景种子节点的扩散,计算出第二阶段显著值;最后,将第二阶段显著值作为中层特征融入图的边权重计算,由此获得新的扩散矩阵,并将整体超像素点作为种子节点进行扩散,计算获得最终显著图。在公开数据集上测试结果表明,改进算法有效提高了查全率、查准率和F-measure等评价指标。

Due to the lack of updating the initial graph model via diffusion results, the traditional diffusion-based manifold ranking model fails to suppress backgrounds and to lose salient objects with small scales on complex data sets. To overcome these problems, the paper proposes an improved saliency detection method via progressive diffusion models. First, we segment an input image via a superpixel segmentation algorithm into superpixels, which are taken as nodes in a graph, and the initial graph is constructed by the traditional manifold ranking algorithm;then the paper diffuses the saliency values via pseudo-background seeds and regard them as the first-step saliency. Second, the paper segments the first-step saliency by adaptive thresholds in order to obtain foreground seeds. Furthermore, the paper diffuses the saliency values via foreground seeds and take them as the second-step saliency. Finally, the paper takes the second-step saliency as mid-level features and fuse them into the calculation of the edge weights, then obtains the new diffusion matrix and takes the whole superpixels as seeds to calculate the final saliency map. Extensive experiments on several public data sets demonstrate that the proposed method improves the precision rate, recall rate and F-measure values.