基于CRF-MR的自顶向下显著性目标检测方法

Top-down saliency target detection dased on CRF-MR

针对自顶向下显著性目标检测边界模糊及准确率低的问题,提出一种结合条件随机场(conditional random field,CRF)和流行排序(manifold ranking,MR)的自顶向下显著性目标检测方法。首先对图像进行超像素分割,以超像素块特征为节点建立无向图;然后输入具有目标先验的CRF中得到节点的显著值,并通过边缘背景先验MR修改显著值;最后扩展初步显著性目标得到最终显著性图。实验结果表明,在行人、汽车和自行车类目标检测中目标边界明确,与基于CRF的方法相比,该方法在保证运算效率的同时具有更好的鲁棒性。

In order to tackle the problems of the target boundary fuzzy and the low accuracy of the top-down saliency target detection method,this paper proposed a top-down significance target detection method combining the conditional random field（CRF）and manifold ranking（MR）.First of all, it super-pixel segmented the image and used the feature factor extracted from the super pixel block as the node to establish the undirected graph model.Then,it got saliency value of the node through the random field with the target prior and modified by the manifold ranking with the edge background prior.Finally,it expanded the initial saliency target to get the final saliency map.Experimental results show that the target boundary is clear in the target detection of pedestrians,cars and bicycles.Compared with the CRF-based method,this method has better robustness while ensuring the efficiency of the operation.