采用SVD协同训练的半监督实例级目标检测  

Semi-supervised instance object detection method based on SVD co-training

作　　者：王睿[1] 樊思杨 许婧文 温志庆 WANG Rui;FAN Siyang;XU Jingwen;WEN Zhiqing(Laboratory of Precision Opto-mechatronics Technology,Ministry of Education,Institute of Instrumentation Science and Opto-electronics Engineering,Beihang University,Beijing 100191,China;Engineering Research Center for Intelligent Robotics,Ji Hua Laboratory,Foshan 528200,China)

机构地区：[1]北京航空航天大学,仪器科学与光电工程学院,精密光机电一体化教育部重点实验室,北京100191 [2]季华实验室,智能机器人工程研究中心,广东佛山528200

出　　处：《光学精密工程》2023年第13期2000-2007,共8页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.61673039);佛山市产业领域科技攻关专项(No.2020001006807)。

摘　　要：在室内实例物体目标检测中,传统深度学习需要大量人工标注的训练样本进行网络训练,费时费力,为此提出并实现了一种采用奇异值分解(Singular Value Decomposition,SVD)和协同训练的半监督实例级目标检测网络SVD-RCNN。挑选关键样本进行人工标注并预训练SVD-RCNN,以确保其获取更多先验知识,采用基于SVD的收敛-分解-微调策略,在SVD-RCNN中得到两个较强独立性的检测器以满足协同训练的要求,最后提出一种自适应的自标注策略,获得高质量的自标注及检测结果。在多个室内实例数据集上对该方法进行测试,在GMU数据集上只需人工标注199个样本,均值平均精度(mean Average Precision,mAP)达到了79.3%,相较于需标注3 851个样本的全监督Faster RCNN的81.3%mAP仅下降了2%。消融实验及系列实验证明了本文方法的有效性和普适性,本文提出的方法仅需人工标注5%的训练数据,即可达到与全监督学习相当的实例级目标检测精度,有利于智能机器人高效识别不同实例物体的实际应用。Detecting indoor instance objects is useful for various applications.Traditional deep-learning methods require a large number of labeled samples for network training,making them time-consuming and labor-intensive.To address this problem,SVD-RCNN—a semi-supervised instance object detection net-work based on singular value decomposition(SVD)and co-training—is proposed.First,key samples are selected for manual labeling to pre-train SVD-RCNN,to ensure that it acquires more prior knowledge.Second,a convergence,decomposition,and finetuning strategy based on SVD is used to obtain two detec-tors with strong independence in SVD-RCNN to satisfy the requirements of co-training.Finally,an adap-tive self-labeling strategy is used to obtain high-quality self-labeling and detection results.The method was tested on multiple indoor instance datasets.On the GMU dataset,it achieved a mean average precision of 79.3%with 199 manually labeled samples.This was only 2%lower than that(81.3%)of Faster RCNN with fully supervised learning,which required labeling 3851 samples.Ablation studies and a series of ex-periments confirmed the effectiveness and universality of the method.The results indicated that the meth-od only needs to manually label 5%of the training data to achieve instance-level detection accuracy compa-rable to that of fully supervised learning;thus,it is suitable for applications in which intelligent robots must efficiently identify different instance objects.

关 键 词：计算机视觉 目标检测 半监督学习 自标注 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]