一种新的鲁棒无监督显著性目标检测方法  

A novel unsupervised robust salient object detection method

作　　者：刘明明 李震霄 郑丽丽 薛雪[3] LIU Mingming;LI Zhenxiao;ZHENG Lili;XUE Xue(School of Mechatronic Engineering,Jiangsu Normal University,Xuzhou,Jiangsu 221008,China;School of Intelligent Manufacturing,Jiangsu Vocational Institute of Architectural Technology,Xuzhou,Jiangsu 221008,China;School of Information and Control Engineering,China University of Mining and Technology,Xuzhou,Jiangsu 221116,China)

机构地区：[1]江苏师范大学机电工程学院,江苏徐州221008 [2]江苏建筑职业技术学院智能制造学院,江苏徐州221008 [3]中国矿业大学信控学院,江苏徐州221116

出　　处：《中国科技论文》2019年第12期1356-1361,共6页China Sciencepaper

基　　金：国家自然科学基金资助项目(61801198);江苏省科技项目(BK20180174)

摘　　要：针对传统无监督图像显著性目标检测鲁棒性不强、学习算法复杂度高的问题,提出了一种新的鲁棒无监督显著性目标检测方法--三元结构化矩阵分解目标检测。该方法利用低秩矩阵三元分解降低奇异值分解(singular value decomposition)的算法复杂度,结合高层先验知识,提升复杂背景下的显著性目标检测性能。通过分层稀疏正则化和构造索引树,解决显著图的细节缺失问题。在3种标准多目标数据集上对主流无监督显著性目标检测方法进行了实验对比,结果表明,所提方法学习时间最多可以降低40%,并且F-measure指标在超过50%的阈值范围内鲁棒性优于当前最好的无监督检测算法。Aiming at the problem of weak robustness and high complexity of learning algorithm in traditional unsupervised image salient object detection, a new robust unsupervised method based on ternary structured matrix decomposition is proposed. This method reduced the algorithm complexity of singular value decomposition by using low rank matrix ternary decomposition, and improved the detection performance of salient objects in complex background by combining high level prior knowledge. Through hierarchical sparse regularization and index tree construction, the problem of missing details of salient images was solved. The experimental comparison of mainstream unsupervised salient object detection method on three standard multi-objective data sets shows that the proposed method can reduce the learning time by up to 40%, and the robustness of F-measure index is better than the current best unsupervised detection algorithm within the threshold range of more than 50%.

关 键 词：图像目标检测 图像显著图生成 无监督鲁棒目标检测 矩阵低秩分解 索引树 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]