融合背景先验与中心先验的显著性目标检测

发布时间：2018-03-07 00:15

本文选题：显著性检测　切入点：中心先验　出处：《中国图象图形学报》2017年05期 　论文类型：期刊论文

【摘要】：目的现有的显著性目标检测算法通常基于单一的先验信息,导致提取的原图像信息不全面,为了解决该问题,提出一种新的基于背景先验与中心先验相融合的显著目标检测算法。方法该方法首先以边缘超像素为吸收节点,利用马尔可夫吸收链计算其他超像素的平均吸收时间作为背景先验值,得到背景先验图;然后使用改进Harris角点检测估计目标区域位置,建立峰值位于目标中心的2维高斯函数,计算各超像素的中心先验值,获取中心先验图;最后将背景先验图与中心先验图相融合得到显著图。同时该方法融合多尺度检测结果,进一步提高显著值的准确性。结果通过ASD、SED1、SED2和SOD 4个公开数据库对比验证了基于背景先验与中心先验相融合算法具有较高的查准率、查全率和F-measure,相较于基于马尔可夫吸收链算法均提高了3%以上,总体效果明显优于目前的10种主流算法。结论相较于基于单一先验信息的算法,基于背景先验与中心先验相融合的算法充分利用了图像信息,在突出全局对比的同时也保留了较多的局部信息,能高亮地凸显图像中的显著性目标。该方法在检测单一目标的图片时效果显著,但对于多目标的显著性检测尚有不足。
[Abstract]:Objective the existing salient target detection algorithms are usually based on a single priori information, which leads to incomplete original image information, in order to solve this problem. A new significant target detection algorithm based on the fusion of background priori and central priori is proposed. Using Markov absorption chain to calculate the average absorption time of other super-pixels as background priori, background prior map is obtained, and then the improved Harris corner is used to detect and estimate the location of the target region, and a two-dimensional Gaussian function with peak value located in the center of the target is established. The center priori value of each superpixel is calculated, and the central priori graph is obtained. Finally, the background priori map is fused with the central priori graph to obtain the salient map. At the same time, the multi-scale detection results are fused by this method. Results the comparison of four open databases, ASD / SED1, SED2 and SOD, showed that the algorithm based on background prior and central prior fusion had high precision. Compared with the Markov absorption chain algorithm, the recall rate and F-measurement are improved by more than 3%, and the overall effect is obviously better than that of the current 10 mainstream algorithms. Conclusion compared with the algorithm based on a single prior information, The algorithm based on the fusion of background priori and central priori makes full use of the image information and keeps more local information while highlighting the global contrast. The method can highlight the salient target in the image. This method is effective in detecting the image of a single target, but it is still insufficient for the significant detection of multiple targets.
【作者单位】：南通大学电气工程学院;德岛大学理工学研究部;
【基金】：江苏省普通高校自然科学研究面上项目(16KJB520037) 国家自然科学基金项目(61602150) 江苏省博士后科研资助计划项目(1601013B) 南通大学研究生科研创新计划项目(YKC15019)~~
【分类号】：TP391.41

【参考文献】