摘要
针对显著性检测方法存在的目标检测不完整的现象,提出一种显著性目标完整性检测方法,可同时考虑目标检测的完整性与显著性.首先,采用分层分割方法获取目标的轮廓生成预处理灰度图.然后,采用自适应阈值分割方法处理基于聚类的显著图,获取超显著性图与超显著性像素点.最后,根据预处理灰度图目标含有超显著性像素点的比例,生成完整后处理灰度图.对超显著性图和完整后处理图进行加权融合,获取最终的完整显著性图.标准数据集上实验仿真结果表明,该方法具有更高的精度与召回率,优于现有的显著性检测方法.
The existing saliency detection methods could only detect the significant part of the target object but cannot completely de- tect the saliency objects. In this work, a complete detection method of saliency objects is proposed. Firstly, hierarchical segmentation approach is utilized to achieve the boundary of all objects and endow them with the initial saliency value in terms of the area and con- nectivity of each object and obtain grayscale image. Then,in order to achieve the super saliency image and superpixels, adaptive thresh- old segmentation approach is chose to deal with the cluster-based salient image. Finally, according to the scale of superpixels of each object in the grayscale image, we achieve corresponding object postprocessing image. The final complete saliency image is obtained by combining complete processing image with super saliency image in a linear summation. The proposed method is illustrated on two pub- lic datasets and compares with other state-of-the-art methods. Extensive experiments demonstrate that either saliency detection in single image or co-saliency detection in multiple images, the proposed schemes to be superior terms of both precision and recall.
出处
《小型微型计算机系统》
CSCD
北大核心
2016年第9期2079-2083,共5页
Journal of Chinese Computer Systems
基金
中央高校基本科研业务费专项资金项目(NS2015092)资助
关键词
显著性目标检测
协同显著性
分层分割
聚类
saliency detection
co-saliency
hierarchical segmentation
clustering
