中心矩形构图先验的显著目标检测

发布时间：2018-05-15 02:31

本文选题：显著目标检测 + 中心矩形构图先验　；参考：《中国图象图形学报》2017年03期

【摘要】：目的许多显著目标检测算法侧重从背景角度进行显著性检测,而从前景角度和空间角度进行显著性检测的算法较少,为了解决这个问题,提出了一种基于中心矩形构图先验的显著目标检测算法。方法假定目标分布在中心矩形构图线附近。首先,对图像进行超像素分割并构造闭环图;其次,提取中心矩形构图线上的超像素特征,并进行流形排序,获取初始显著值;然后,通过基于中心矩形构图线获取的初始显著值确定中心矩形构图交点显著值和紧凑性关系显著值;最后,融合三者获得最终的中心矩形构图先验显著图。结果通过MSRA-1000,CSSD,ECSSD,THUS-10000数据集对比验证了中心矩形构图先验算法有较高的准确度和最高的F-measure值,整体效果上优于目前先进的几种算法。且处理单幅图像的平均时间为0.673 s,相比与其他算法也有较大优势。结论从前景角度和空间角度考虑的中心矩形构图先验的显著目标检测算法相比于传统的算法更加具有鲁棒性,无论图像是复杂的还是简单的,都取得很好的检测效果,充分说明算法的有效性。
[Abstract]:Objective in order to solve this problem, many salient target detection algorithms focus on significance detection from background point of view, but less from foreground and spatial perspective. A significant target detection algorithm based on a priori center rectangle composition is proposed. Methods it is assumed that the target is distributed near the center rectangular composition line. Firstly, the image is partitioned by super-pixel and the closed loop graph is constructed. Secondly, the super-pixel feature on the center rectangular composition line is extracted, and the manifold sorting is carried out to obtain the initial significant value. Based on the initial significant value obtained from the central rectangular composition line, the intersection significant value and the compactness significant value of the central rectangular composition are determined. Finally, the final priori significant map of the central rectangular composition is obtained by fusion of the three. Results by comparing the data sets of MSRA-1000 / CSSDX ECSSD-THUS-10000, it is proved that the priori algorithm of center rectangular composition has higher accuracy and highest F-measure value, and the overall effect is better than that of the current advanced algorithms. The average processing time of a single image is 0.673 s, which is also superior to other algorithms. Conclusion the salient target detection algorithm based on a prior centroid rectangle composition is more robust than the traditional algorithm from the perspective of foreground and space. The image is complex or simple, and the detection effect is very good. The effectiveness of the algorithm is fully illustrated.
【作者单位】：安徽大学计算机科学与技术学院;
【基金】：国家科技支撑计划基金项目(2015BAK24B00) 高等学校博士学科点专项科研基金联合项目(20133401110009) 安徽高校省级自然科学研究项目(KJ2015A009)~~
【分类号】：TP391.41

【相似文献】