基于扩散的图像显著性检测

Diffusion-Based Image Saliency Detection

显著性扩散方法是众多显著性检测方法中的一种,此过程中关键是如何更好地选择种子点并进行有效传播。当图像背景比较复杂尤其是图像的前景和背景颜色接近时,种子点的选择会出现错误,从而影响最终显著图的质量。从显著性图出发提出一种能够准确找到显著性种子点的方法,并将随机游走方法改进和流形排序方式融合进行显著性扩散。实验结果表明该方法较之前方法有了很大的改进,得到了更加精细的显著性图。

Diffusion-based saliency detection is one of many methods about image saliency detection.The key is how to choose these seeds and propagate their saliency.Given the unconspicuous difference of color and contrast between the background and the salient object,the erroneous seeds may be selected.Finally,these wrong seeds could affect the quality of saliency map.In this paper,we propose a novel seeds detection approach that take advantage of saliency map.By taking regularized random walk and manifold ranking into account,the saliency map of an image is then estimated.Experimental result demonstrates the proposed method performs well and provide pixel-wised saliency map.