多尺度构图先验的显著目标检测

Multi-scale saliency detection based on composition prior

目的针对基于对比度的显著检测方法,因忽略了特征的空间分布而导致准确性不高的问题,启发于边界先验关于图像空间布局的思想,提出构图先验的显著检测方法。方法假定目标分布于三分构图线周围,根据相关性比较计算显著值。首先,对图像进行多尺度超像素分割并构造闭环图;其次,提取构图线区域超像素特征并使用Manifold Ranking算法计算显著目标与背景的分布;然后,从目标和背景两个角度对显著值进行细化并利用像素区别性对像素点的显著值进行矫正;最后,融合多尺度显著值得到最终显著图。结果在公开的MSRA-1000、CSSD、ECSSD数据集上验证本文方法并与其他算法进行对比。本文方法在各数据集上准确率最高,分别为92.6%,89.2%,76.6%。且处理单幅图像平均时间为0.692 s,和其他算法相比也有一定优势。结论人眼视觉倾向于在构图线周围寻找显著目标,构图先验是根据人眼注意机制研究显著性,具有合理性,且构图先验的方法提高了显著目标检测的准确性。

Objective Saliency detection is a fundamental part of computer vision applications, the goal is to detect impor- tant pixels or regions in an image which attracts human visual attention most. Recently, people have proposed boundary prior, or background information to enhance saliency detection. Such methods even achieve state-of-the-art result, sug- gesting that boundary prior is effective. Compared with most existing bottom-up methods which consider saliency based on the contrast between salient objects and their surrounding regions, boundary prior characterizes the spatial layout of image regions with respect to image boundaries. Inspired by this idea, we propose image composition prior to detect saliency. Observing from images, we find salient objects usually placed in center regions while background lies in boundary re- gions. And images are usually formed with some composition rules, such as Rule of Thirds. Method We propose compo- sition prior method by assuming objects are distributed near composition lines. We select regions near composition lines as initial seeds, and compute saliency according to feature relevance. To be specific, firstly, we segment the image into multi scales and construct a close-loop graph where each node is a super pixel. Secondly, we use nodes which near com- position lines as queries, and extract their features to rank the relevancies of all the other regions by Manifold Ranking, and then compute saliency based on the ranking result. Thirdly, according to the last step, we iteratively refine saliency in the perspective of both object and background. Then assign the saliency value to each pixel. Considering the distinct- ness of different pixels in the same region, we need to correct their saliency. We choose to add a correction value to each pixel based on their distance to feature center. Finally, the saliency detection is carried out by integrating multi- scale saliency. Result In comparison experiments on datasets of MSRA-1000, CSSD, and ECSSD, our method per- forms well when against the state-of-the-art methods. It gets highest precision on three datasets （92.6% , 89.2% , and 76.6% respectively） . The average run time of a single image is 0. 692, which still has some advantages com- pared with other algorithms. Conclusion This study presents a new salient detection method based on composition pri- or. Human vision has the tendency of detecting saliency from regions near composition lines rather than image bounda- ries. Composition prior detects saliency based on human vision mechanism. Experimental results demonstrate detect saliency in the perspective of image composition is reasonable, and using composition prior can improve the detecting accuracy.