基于定位置信度加权的半监督目标检测算法  

Semi-Supervised Object Detection Algorithm Based on Localization ConfidenceWeighting

作　　者：冯泽恒 王丰 FENG Zeheng;WANG Feng(School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学信息工程学院,广州510006

出　　处：《计算机工程与应用》2024年第6期249-258,共10页Computer Engineering and Applications

基　　金：国家自然科学基金(61901124);广东省自然科学基金(2021A1515012305);广州市基础研究计划项目(202102020856)。

摘　　要：为解决伪标签筛选过程的位置噪声数据问题,提出了基于定位置信度加权的Soft Teacher-LAH半监督目标检测算法。通过离散化目标检测网络定位分支的预测输出,引入具有定位感知功能的输出结构LAH。基于LAH预测输出,定义一种衡量定位精度的置信度指标,设计基于该置信度加权的无监督定位损失函数,降低伪标签位置噪声对模型训练的负面影响。实验结果表明了该算法的性能优势,针对微软COCO数据集,在有标注数据占比训练集分别为1%、5%和10%的场景下,该算法相比于现有Soft Teacher方案的平均精度分别提高了1.1、1.2和1.5个百分点;针对PASCAL VOC数据集,在使用VOC07和VOC12分别作为有标注和无标注训练数据的场景下,该算法相比Soft Teacher方案的平均精度提高了1.6个百分点。To address the problem of location noise data in the process of screening pseudo-labels,a semi-supervised object detection algorithm based on localization confidence weighting named Soft Teacher-LAH is proposed.Firstly,it introduces a localization-aware output structure LAH by discretizing the network output of localization branch in object detection model.Secondly,a certain confidence index is defined to measure the localization accuracy based on the predic-tion output of LAH,and an unsupervised localization loss function based on the confidence weighting is designed,which can reduce the negative effect of location noise of pseudo-labels on model training.Experimental results show the perfor-mance advantage of the proposed algorithm,for MS COCO datasets,the average accuracy of the proposed algorithm is improved by 1.1,1.2 and 1.5 percentage points compared with the existing Soft Teacher scheme when the proportion of labeled data in the training set is 1%,5%and 10%respectively.For PASCAL VOC dataset,the average accuracy of the proposed algorithm is improved by 1.6 percentage points compared with Soft Teacher scheme,when VOC07 and VOC12 are used as labeled and unlabeled training data respectively.

关 键 词：目标检测 半监督学习 伪标签 位置噪声 定位置信度 定位损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]