基于序贯三支决策的半监督目标检测算法  

Semi-supervised Object Detection with Sequential Three-way Decision

作　　者：宋法兴 苗夺谦[1,2] 张红云[1,2] SONG Faxing;MIAO Duoqian;ZHANG Hongyun(College of Electronic and Information Engineering,Tongji University,Shanghai 201804,China;Key Laboratory of Embedded System and Service Computing,Ministry of Education,Shanghai 201804,China)

机构地区：[1]同济大学电子与信息工程学院,上海201804 [2]嵌入式系统与服务计算教育部重点实验室,上海201804

出　　处：《计算机科学》2023年第10期1-6,共6页Computer Science

基　　金：国家重点研发计划(2022YFB3104700);国家自然科学基金(61976158,61976160,62076182,62163016,62006172);江西省自然科学基金重点项目(20212ACB202001);江西省“双千计划”。

摘　　要：深度学习对大规模数据的需求以及目标检测标注任务的复杂性促进了半监督目标检测任务的发展。近年来,半监督目标检测已经取得了很多优秀的成果。然而,伪标签中的不确定性依然是半监督目标检测研究中难以避免的问题,优越的半监督方法要求选取合适的过滤阈值来权衡伪标签的噪声信息比例和召回率,以最大程度保留准确有效的伪标签。为了解决此问题,在半监督检测的框架中引入了序贯三支决策算法,将模型输出的伪标签根据不同的筛选阈值划分为干净的前景标签、有噪声的前景标签,以及干净的背景标签,并对其采取不同的处理策略。对有噪声的前景标签采用负类学习损失来学习这些存在噪声的标签,避免学习到其中的噪声信息。实验结果表明了所提算法的性能优势,针对COCO数据集,在有监督数据占比只有10%的情况下,该方法实现了35.2%的检测精度,相比仅依靠有监督训练性能提升了11.34%。The need for large scale data in deep learning and the complexity of object detection annotation task promote the deve-lopment of semi-supervised object detection.In recent years,semi-supervised object detection has achieved many excellent results.However,the uncertainty in pseudo labels is still an unavoidable problem in semi-supervised object detection.The superior semi-supervised method requires an appropriate filtering threshold to balance the proportion of pseudo labels’noise and the recall rate,so as to retain accurate and effective labels as much as possible.To solve this problem,this paper introduces a sequential three-way decision algorithm into semi-supervised object detection,which divides the model output pseudo-labels into clean foreground labels,noisy foreground labels,and clean background labels according to different filtering thresholds,and adopts different processing strategies for them.For noisy foreground labels,we use negative class learning loss to learn these noisy labels,thereby avoiding learning noise information from them.Experimental results show the performance advantage of this algorithm.For COCO dataset,this method achieves performance of 35.2%when supervised data only accounts for 10%,which outperforms the supervised results by 11.34%.

关 键 词：序贯三支决策 不确定性 负类学习 半监督学习 半监督目标检测 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]