基于区域合并的图像显著性检测  

Image Saliency Detection Based on Region Merging

作　　者：孙丰[1,2] 秦开怀[1] 孙伟 郭华源[1] 

机构地区：[1]清华大学计算机科学与技术系,北京100084 [2]中国卫星海上测控部试验技术部,江阴214431 [3]陆航研究所航电研究室,北京100021

出　　处：《计算机辅助设计与图形学学报》2016年第10期1679-1687,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：教育部博士点基金(20130002110002)

摘　　要：为了提高自然图像显著性检测准确度,提出一种基于区域合并的图像显著性检测算法.该算法直接通过区域合并的方式逐渐将图像从初始状态下的多个区域合并为显著性对象和背景2个区域.在合并过程的不同阶段采用了不同的合并策略,首先利用超像素分割方法将图像分为若干初始区域,在第一阶段仅合并相似且相邻的区域,使得属于同一对象的像素合并到同一区域;然后,处理上述过程中产生的空洞以及因遮挡造成的属于同一对象的区域不相邻的情况;再在区域显著分析的引导下,不断将显著性最差的区域合并到背景区域,而不是尝试将显著性区域合并到一起.最后利用合并过程中得到的多个候选显著性区域加权得到最终的显著性区域结果.在2个公开测试集上进行了测试并与其他算法进行了对比,实验结果证明了文中算法的有效性;特别是在难度更大的ECSSD数据集中,该算法的准确度要优于同类算法.In this paper, an image saliency detection algorithm based on region merging is proposed. Our goal is to divide the image into the background region and the salient region directly via a region merging process. To achieve this, different merging strategies are employed in different stages. Firstly, the image is over-segmented into superpixels. In the first stage, only similar neighboring superpixels are allowed to be merged together. Then, in the second stage, occlusions and hole-regions produced in the last step are handled. At last, instead of merging salient regions together, non-salient regions are merged into the background region guided by the region saliency analysis. The final saliency result can be obtained via a weighted average of several saliency region proposals obtained during the region merging process. Our algorithm is tested on the two public datasets and compared with other state-of-the-art algorithms. The experimental results show that our algorithm is effective and efficient. Particularly, for the more challenging dataset ECSSD, the accuracy of our algorithm outperforms other region-based state-of-the-art algorithms.

关 键 词：显著性检测 区域合并 超像素 图像分割 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]