基于背景感知的显著性目标检测算法

摘要

在显著性目标检测算法中, 流形排序的检测方法存在先验背景假设和目标检测不完整的问题. 针对该问题,在流形排序算法基础上, 融入背景鉴别、BING特征估计和权重调整, 提出了一种基于背景感知的显著性目标检测算法. 首先, 通过计算颜色聚类后的边界区域的综合差异度, 得到真实背景种子点, 从而感知到真实背景区域; 再结合图像的BING特征与初始显著图信息, 获取目标位置, 从而得到完整的前景种子点区域; 然后重构前景区域的图模型且利用加权k-壳分解法, 来调整前景区域节点之间的连接权重, 进而获得清晰的目标边界. 实验结果表明, 同当前经典的一些算法比较, 本文算法在准确率、召回率、F-measure和平均MAE上都优于其余算法.%In the saliency detection algorithm, there are some problems in the detection of the manifold ranking, such as the over ideal of the background and the incomplete target detection. Aiming at these problems, this study incorporated background identification, BING feature estimation, and weight adjustment in traditional manifold rank algorithm, and a method was proposed based on background awareness. Firstly, through the adaptive color clustering of the boundary area and calculating the synthetic difference degree to get the real background seed point, the real background areas were sensed. Then the BING feature of the image was calculated and the saliency map information was combined to obtain the target position, so as to obtain the complete foreground seed point area. Next, by reconstructing the graph model of the foreground region and using the weighted k-shell decomposition method, we adjusted the connection weight between the nodes in the foreground region to obtain a clear target boundary. The experimental results show that the proposed algorithm is superior to other algorithms in terms of precision, recall, F-measure, and average MAE compared with some classical algorithms.