摘要
针对城市场景标签获取的高额成本问题,文中提出结合自集成和对抗学习的域自适应城市场景语义分割方法.对于源域和目标域的较大域间差异问题,采用风格转换的方法将源域数据集合成具有目标域风格的新数据集,作为新的源域数据集,从而有效减少源域与目标域的域间差异.对于目标域的域内差异问题,引入自集成方法,构造教师网络,利用教师网络在目标域分割图上通过一致性约束监督与指导学生网络,从而减小目标域的域内差异,提高分割精度.采用自训练的方法获得目标域的伪标签,将伪标签加入对抗学习方法中,重新训练网络模型,进一步提高模型的分割能力.在数据集上的分割实验表明文中方法的有效性.
Aiming at the problem of high cost of urban scene label acquisition,an algorithm of domain adaptation semantic segmentation for urban scene combining self-ensembling and adversarial learning is proposed.For the inter-domain gap between source and target domains,the method of style transfer is employed to transfer the source domain into a new dataset with the style of target domain.For the problem of intra-domain gap in the target domain,the self-ensembling method is introduced and a teacher network is constructed.The teacher network is utilized to supervise and guide the student network through consistency constraints on the target domain segmentation map to reduce the intra-domain gap of the target domain and improve the segmentation accuracy.The self-training method is exploited to obtain the pseudo label of the target domain and add the pseudo label into the adversarial learning method to retrain the network and further improve the segmentation ability.Experiments on segmentation datasets verify the effectiveness of the proposed algorithm.
作者
张桂梅
鲁飞飞
龙邦耀
缪君
ZHANG Guimei;LU Feifei;LONG Bangyao;MIAO Jun(Institute of Computer Vision,Nanchang Hangkong University,Nanchang 330063)
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2021年第1期58-67,共10页
Pattern Recognition and Artificial Intelligence
基金
国家自然科学基金项目(No.61462065,61661036)资助。
关键词
自集成
对抗学习
域自适应
城市场景
语义分割
Self-ensembling
Adversarial Learning
Domain Adaptation
Urban Scene
Semantic Segmentation
作者简介
通信作者:张桂梅,博士,教授,主要研究方向为计算机视觉、图像处理、模式识别.E-mail:guimei.zh@163.com;鲁飞飞,硕士研究生,主要研究方向为计算机视觉、图像处理、模式识别.E-mail:1323568545@qq.com;龙邦耀,硕士研究生,主要研究方向为计算机视觉、图像处理、模式识别.E-mail:lby609527215@163.com;缪君,博士,教授,主要研究方向为基于图像的三维重建、图像处理、模式识别.E-mail:miaojun@nchu.edu.cn.
