基于多尺度掩码分类域自适应网络的跨域目标检测算法  

Cross-Domain Object Detection Algorithm Based on Multi-scale Mask Classification Domain Adaptive Network

作　　者：胡杰[1,2,3] 徐博远 熊宗权 昌敏杰 郭迪[1,2,3] 谢礼浩 Hu Jie;Xu Boyuan;Xiong Zongquan;Chang Minjie;Guo Di;Xie Lihao(Wuhan University of Technology,Hubei Key Laboratory of Advanced Technology for Automotive Components,Wuhan 430070;Wuhan University of Technology,Hubei Collaborative Innovation Center for Automotive Components Technology,Wuhan 430070;Wuhan University of Technology,Hubei Research Center for New Energy,Intelligent Connected Vehicle,Wuhan 430070)

机构地区：[1]武汉理工大学,现代汽车零部件技术湖北省重点实验室,武汉430070 [2]武汉理工大学,汽车零部件技术湖北省协同创新中心,武汉430070 [3]武汉理工大学,湖北省新能源与智能网联车工程技术研究中心,武汉430070

出　　处：《汽车工程》2022年第9期1327-1338,共12页Automotive Engineering

基　　金：湖北省科技重大专项(2020AA001)资助。

摘　　要：针对无监督域自适应目标检测中,域的可辨性和不变性之间的矛盾导致域负迁移和多尺度问题,本文中提出了一种可缓解域负迁移的多尺度掩码分类域自适应网络。首先在主干网络上对多个中间层进行图像级域对抗训练。接着在图像级特征图上加入区域提议掩码,作为一种补充信息对实例特征进行补充。最后提出分类别实例级域分类器,在保证域可辨性前提下,使网络尽可能地提取出有效的域不变信息。在Cityscapes和FoggyCityscapes两个数据集上进行验证的结果表明,本文提出的多尺度掩码分类域自适应网络,其域分类平均精度的平均值提高了13.2个百分点,说明网络域自适应能力显著提升。Aiming at the problems of multi-scale and domain negative transfer caused by the contradiction between domain discriminability and invariance in unsupervised domain adaptive object detection,a multi-scale mask classification domain adaptive network(MMCN),that can alleviate the negative transfer of domain,is proposed in this paper.Firstly the adversarial training of image-level domain is performed on multiple intermediate layers on backbone network.Then a region proposal mask is added to the image-level feature map as a supplementary information to supplement instance features.Finally a sub-category instance-level domain classifier is put forward to enable the network to extract effective domain-invariant information as much as possible on the premise of ensuring domain discriminability.The results of verification on both Cityscapes and FoggyCityscapes datasets show that the mean average precision of domain classification with MMCN proposed is 13.2 percentage points higher than that with DA-FasterRCNN,significantly enhancing the domain adaptive capability of network.

关 键 词：目标检测 域自适应网络 深度学习 对抗学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]