融合颜色显著性和似物性的显著物体检测

Object-level saliency fusing with color saliency and objectness

文中提出一种融合颜色显著性和似物性的对象级显著性检测方法.首先量化并计算像素点在RGB空间下的稀疏性以及在LAB空间中的对比度,得到像素级的显著性;其次对原图像进行超像素分割,以超像素内所有像素的显著性均值作为该超像素的显著值,通过能量函数平滑优化相邻的超像素,得到前景分割明显且变化较小的超像素级显著性;然后优化原似物性方法,通过中心环绕法则融合显著性值和似物性值,作为对象级显著性的贝叶斯先验概率,计算似物性窗口内显著物体与窗口像素的比值,得到似然概率;最后通过贝叶斯模型计算,得到显著物体后验概率,并以此作为似物性窗口的评分标准.在MSRA-1000和ECSSD数据集上先将颜色显著性方法与8种state-of-the-art显著性方法相比较,显著性检测效果较好,检测精度高于其他几种方法.将优化后的似物性与另外3种方法对比,优化结果能使检测窗口大量覆盖于显著图中目标物体上方.将文中显著物体检测方法与同样融合了似物性的显著物体检测方法进行比较,结果表明文中方法较其他方法能够更精确、更完整检测出显著物体.

We propose a novel object-level saliency method which fuses color saliency with objectness.Firstly,pixel-level saliency is obtained by computing the pixel’s sparseness in RGB space and the contrast in LAB space.The image is segmented into super-pixels and the average saliency of all the pixels within a super-pixel is considered as its saliency value.We calculate super-pixel saliency by averaging pixel-level saliency value which is in each super-pixel.An energy function is used to smooth the saliency value of adjacent super-pixels,which is helpful to obtain a smooth saliency map with distinct foreground object.Secondly,the classic objectness method is optimized and the obtained objectness is fused with saliency map through the center-surround principal.The fusion result is considered as the prior probability of object-level saliency.The likelihood probability of object-level saliency is obtained by calculating the ratio between the foreground cluster in the objectness window and the total pixels in the window.Finally,an object-level saliency posterior probability is calculated by the Bayesian model.Saliency experiments on MSRA-1000 and ECSSD dataset have shown that our saliency method have a better performance and high detection precision than the other eight state-of-the-art algorithms.We also experiment the performance between ours with other three proposal measure,and the proposal windows generated by our method largely cover the foreground object in the color saliency map.Finally,object-level saliency experiments on the dataset show that our method has a better performance than the other similar methods which detect salient object with objectness.