基于自适应背景模板与空间先验的显著性物体检测方法

发布时间：2018-02-22 14:21

本文关键词： 显著性检测背景模板传播机制空间先验　出处：《自动化学报》2017年10期 　论文类型：期刊论文

【摘要】：目前,显著性检测已成为国内外计算机视觉领域研究的一个热点,但现有的显著性检测算法大多无法有效检测出位于图像边缘的显著性物体.针对这一问题,本文提出了基于自适应背景模板与空间先验的显著性物体检测方法,共包含三个步骤:第一,根据显著性物体在颜色空间上具有稀有性,获取基于自适应背景模板的显著图.将图像分割为超像素块,提取原图的四周边界作为原始背景区域.利用设计的自适应背景选择策略移除原始背景区域中显著的超像素块,获取自适应背景模板.通过计算每个超像素块与自适应背景模板的相异度获取基于自适应背景模板的显著图.并采用基于K-means的传播机制对获取的显著图进行一致性优化;第二,根据显著性物体在空间分布上具有聚集性,利用基于目标中心优先与背景模板抑制的空间先验方法获得空间先验显著图.第三,将获得的两种显著图进行融合得到最终的显著图.在公开数据集MSRA-1000、SOD、ECSSD和新建复杂数据集CBD上进行实验验证,结果证明本文方法能够准确有效地检测出图像中的显著性物体.
[Abstract]:At present, salience detection has become a hot topic in the field of computer vision at home and abroad, but most of the existing salience detection algorithms can not effectively detect salient objects at the edge of images. In this paper, a significant object detection method based on adaptive background template and spatial priori is proposed. It consists of three steps: first, according to the rarity of salient object in color space, A salient image based on adaptive background template is obtained. The image is divided into super-pixel blocks, and the edges of the original image are extracted as the original background area. The proposed adaptive background selection strategy is used to remove the significant super-pixel blocks in the original background region. The adaptive background template is obtained. The saliency map based on adaptive background template is obtained by calculating the difference between each super-pixel block and the adaptive background template. The K-means based propagation mechanism is used to optimize the coherence of the obtained salient image. According to the clustering of salient objects in spatial distribution, a spatial priori map is obtained by using a spatial priori method based on target center priority and background template suppression. Finally, the final salience map is obtained by fusion of the two salient maps. The experimental results on the open data set MSRA-1000 SOD ECSSD and the newly built complex dataset CBD show that the proposed method can accurately and effectively detect the salient objects in the image.
【作者单位】：南京航空航天大学计算机科学与技术学院;软件新技术与产业化协同创新中心;江苏省委党校;
【基金】：中央高校基本科研业务费专项资金(NS2015092)资助~~
【分类号】：TP391.41

【相似文献】