面向域自适应目标检测的一致无偏教师模型  

Consistent and unbiased teacher model research for domain adaptive object detection

作　　者：程旭东 史彩娟 高炜翔 王森 段昌钰 闫晓东 CHENG Xudong;SHI Caijuan;GAO Weixiang;WANG Sen;DUAN Changyu;YAN Xiaodong(College of Artificial Intelligence,North China University of Science and Technology,Tangshan Hebei 063210,China;Hebei Key Laboratory of Industrial Intelligent Perception,Tangshan Hebei 063210,China)

机构地区：[1]华北理工大学人工智能学院,河北唐山063210 [2]河北省工业智能感知重点实验室,河北唐山063210

出　　处：《图学学报》2025年第1期114-125,共12页Journal of Graphics

基　　金：北京市现代信息科学与网络技术重点实验室开放课题(XDXX2301);华北理工大学杰出青年基金(JQ201715);唐山市人才项目(A202110011)。

摘　　要：作为一种重要手段,自训练方法极大提升了域自适应目标检测(DAOD)性能,其主要通过教师网络对目标域数据进行预测,然后选择高置信度的预测结果作为伪标签来指导学生网络训练。然而,由于源域与目标域存在显著的域差异,教师网络产生的伪标签质量不佳,进而影响学生网络训练,降低了模型性能。因此,提出一种面向DAOD的一致无偏教师(CUT)模型。首先,在教师网络设计自适应阈值生成(ATG)模块,该模块通过高斯混合模型(GMM)在训练过程为每张图像生成自适应阈值筛选伪标签,保证伪标签数量时序一致性,提高伪标签质量。其次,提出预测引导样本选择(PSS)策略,借助教师网络中区域建议网络的预测结果为学生网络选择样本,使选择的样本与真实结果具有一致性,降低质量不佳伪标签对学生网络的影响。此外,为了提升对小目标和数量较少困难类别目标的检测性能,设计混合域增强(MDA)模块,在训练过程中生成包含源域和类目标域随机信息的混合域图像对学生网络进行训练。将该模型在3个场景数据集进行实验,性能分别提升4.0%,5.8%和3.7%,验证了该算法的有效性。值得注意的是,该模型CUT首次利用自训练方法来解决可见光图像到红外图像的较大域差异问题。As a significant approach,the self-training method has significantly enhanced the performance of domain adaptive object detection methods.The self-training method primarily predicts target domain data through a teacher network,and then selects high-confidence predictions as pseudo-labels to guide student network training.However,due to significant domain differences between the source and target domains,the pseudo-labels generated by the teacher network are often of poor quality,adversely impacting student network training and reducing the performance of object detection.To address this challenge,a consistent and unbiased teacher(CUT)model for domain adaptive object detection was proposed.Firstly,an adaptive threshold generation(ATG)module was designed within the teacher network.The ATG module utilized a Gaussian mixture model(GMM)during training to generate adaptive thresholds for each image,ensuring temporal consistency of pseudo-label quantities and thereby enhancing their quality.Secondly,a prediction-guided sample selection(PSS)strategy was introduced,which leveraged predictions from the region proposal network within the teacher network to guide the selection of positive and negative samples for the student network.The PSS strategy effectively aligned the selected samples with real outcomes,thereby mitigating the impact of low-quality pseudo-labels on the student network.Furthermore,to improve detection performance for small objects and challenging objects with fewer instances,a mixed domain augmentation(MDA)module was devised to generate mixed-domain images containing random information from both the source and target-like domains to supervise student network training.Extensive experiments conducted on three scenario datasets demonstrated the effectiveness of the proposed CUT,with performance improvements of 4.0%,5.8%,and 3.7%,respectively.Notably,the proposed CUT model applied the self-training method for the first time to address the problem of large domain disparities between visual images and infrared

关 键 词：域自适应 目标检测 自训练 伪标签 一致性 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]