基于贝叶斯框架的显著物体检测

Salient Object Detection Based on Bayesian Framework

传统显著性方法仅能检测出目标物体的边缘或角点等方面信息,无法完整分割出显著物体整体。为解决以上问题,论文考虑以超像素为基本处理单元,而不是仅仅计算单个像素点信息,通过综合考虑相邻超像素之间的紧致性及颜色关系等因素,优化得到超像素级显著图。其次,通过似物性计算目标物体粗糙检测区域作为位置信息。最后通过贝叶斯框架融合各部分先验信息,计算得到最终的显著物体后验概率,并以此为优化后的似物性方法评分标准,最终检测出显著物体。在标准公开数据集MSRA上的实验结果表明论文超像素级显著图精度明显高于其他8种传统显著性算法,最终在显著图上的搜索能够完整检测出显著物体。

Traditional saliency measure can only detect the margin or corner point of the target object.They almost can't seg?ment the whole body of the object from the image.To solve the problem,the paper proposes a novel salient object detection measure calculated by super-pixel.Firstly,it calculates the pixel level saliency map and obtains the super-pixel level saliency map by aver?aging the pixel level saliency value among the super-pixel.The final super-pixel level saliency map is obtained through lagrange function optimization.Secondly,the object location information is gotten by objectness measure.Finally,the saliency object is de?tected by Bayesian posterior probability.Experiments on MSTA datasets show that our saliency measure has a better detection perfor?mance and high precision than the other 8 state-of-the-art saliency measure.And also our salient object detection measure has a better detection performance than the generic objectness measure.