为了解决目前主流的显著性检测算法在复杂多目标遥感图像中检测能力不足的问题, 提出一种基于超像素区域相似性度量的显著目标提取算法。该算法利用简单线性迭代聚类方法对原始图像进行超像素分割, 通过基于图论的视觉显著性方法检测出显著超像素, 并对其修正得到显著目标提取的训练样本, 进一步逐层计算全体超像素区域与显著超像素区域的相似性并转化为超像素区域的隶属度值, 最后实现对整幅超像素图像的显著目标提取。实验结果表明, 该算法具有较高的准确率和召回率, 能更加有效地检测出遥感图像中的显著目标, 提取效果优于主流的显著性检测算法, 还可以有效应用于复杂多目标的遥感图像显著目标信息提取中。

In order to optimize the insufficient ability for complex multi-target remote sensing image detection using the reported saliency algorithms, an extraction algorithm of salient object based on similarity measurement for superpixel regions is proposed. The original image is segmented into certain superpixel regions using simple linear iterative clustering method, and some high saliency regions are extracted correctly using graph-based visual saliency method. Meanwhile, parts of the edge superpixels need to be amended and the rest of salient superpixels are used as training samples. By calculating the similarity of all superpixels and training samples hierarchically, a reasonable membership value of each superpixel is established to separate the goal superpixel regions with high saliency. Finally, all the superpixels salient objects from the original images are extracted successfully using the membership values. The experimental results show that the proposed algorithm has higher precision and recall rates than the other saliency detection methods, thus it can be effectively applied to complicated multi-objective in target information extraction of remote sensing images significantly.