改进中心先验的显著性检测

Saliency detection via improved center prior

针对传统基于中心先验的显著性检测算法不能均匀突出显著性目标的问题,提出改进中心先验的显著性检测算法.首先将输入图像分割成4种不同超像素尺度,对输入图像计算凸包;然后计算4种不同尺度下的背景显著图和前景显著图,将对应尺度下的背景显著图和前景显著图融合得到该尺度下的融合显著图,将凸包和4种不同尺度的融合显著图作为先验,计算4种不同尺度下的中心先验图;最后将得到的不同尺度的中心先验图经MKB算法优化得到最终显著图.在MSRA1000,PASCAL和ECSSD数据集上进行,实验结果表明所述算法在显著目标检测的准确性方面优于14种对比算法.

Aiming at the problem that traditional saliency detection algorithms based on center priors cannot uniformly highlight salient object,this paper proposes an improved center prior saliency detection.First,the image is divided into superpixels of four different scales and the convex hull of the image is calculated.Then,the background saliency map and the foreground saliency map at four different scales are calculated.Third,the background saliency map and foreground saliency map at the corresponding scale are fused to obtain the fused saliency map.Fourth,the convex hull and the fusion saliency map of four different scales are used as prior information to calculate the center prior map at four different scales.Finally,the center prior maps of different scales are optimized by MKB algorithm to obtain the final saliency map.Tests based on three types of public data sets,MSRA1000,PASCAL and ECSSD show that compared with the current 14 state-of-the-art algorithms,the algorithm performs well in terms of accuracy.