融合轮廓提取的显著性物体完整性检测方法-《计算机技术与发展》

文章信息/Info

Title:: A Complete Salient Object Detection Method Based on Contour Prior and Background Prior

作者:: 李巍¹ ; 王鸥¹ ; 田庆阳¹ ; 梁凯¹ ; 谭学军¹ ; 刚毅凝¹ ; 刘嘉华² ; 林华锋³; 1. 国网辽宁省电力有限公司,辽宁沈阳 110004;2. 南瑞集团有限公司(国网电力科学研究院有限公司),江苏南京 211000;3. 南京航空航天大学计算机学院,江苏南京 211106

Author(s):: LI Wei¹ ; WANG Ou¹ ; TIAN Qing-yang¹ ; LIANG Kai¹ ; TAN Xue-jun¹ ; GANG Yi-ning¹; LIU Jia-hua² ; LIN Hua-feng³; 1. State Grid Liaoning Electric Power Supply Co. ,Ltd. ,Shenyang 110004,China;2. Nari Group Corporation/ State Grid Electric Power Research Institute,Nanjing 211000,China;3. Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China

Keywords:: salient object detection; background template; contour extraction; superpixel segmentation

摘要:: 针对现有基于背景模板的方法不能完整检测部分区域与背景具有相似特征的显著性目标这一问题,提出了一种基于轮廓检测的显著性目标完整性检测方法。首先,提取输入图像的初始轮廓,利用提出的基于虚拟连接的轮廓处理方案合并相邻的轮廓并移除孤立的轮廓,利用设计的基于最短路径的闭环搜寻方案合并距离较远的轮廓,获取基于轮廓检测的显著图。然后,利用自适应阈值分割算法处理基于背景模板抑制的显著图,获取二值化显著图与显著像素点。通过去除含有显著像素点比例小于指定阈值的轮廓完整区域,获取优化的基于轮廓检测的显著图。最后,将其与二值化显著图进行融合,获取完整显著图。实验结果表明,该方法针对显著性目标位于任意位置的图像均能取得较好的显著图。

Abstract:: Aiming at the problem that the existing methods cannot completely detect the salient object whose regions have the similar features with background template,we propose a complete detection method for salient object based on contour detection. Firstly,the initial contour of the input image is extracted. The proposed contour processing scheme based on virtual link is used to merge the adjacent contour and remove the isolated contour. The designed closed-loop search scheme based on shortest path is used to merge the contour with a long distance and obtain the salient map based on contour detection. Secondly,adaptive threshold segmentation algorithm is used to process the salient map based on background template suppression,and binarization salient map and salient pixel points are obtained. Thirdly,the optimized contour detection-based salient map is obtained by removing the contour intact region with the proportion of significant pixel points less than the specified threshold. Finally,the complete salient objects are detected by fusing the salient map based on contour detection and background template based salient map. Experiment demonstrates that the proposed method outperforms other state-of-the-art approaches and completely detects the salient object which locates at random positions.