第17卷第4期 智能系统学报 Vol.17 No.4 2022年7月 CAAI Transactions on Intelligent Systems Jul.2022 D0:10.11992/tis.202106020 网络出版地址：https:/ns.cnki.net/kcms/detail/23.1538.TP.20220608.1352.002.html 双层残差语义分割网络及交通场景应用 谭睿俊，赵志诚，谢新林 (太原科技大学电子信息工程学院，山西太原00024) 摘要：针对图像语义分割过程中特征提取网络的深度问题以及下采样池化层降低特征图分辨率等问题，提出 了一种基于双层残差网络特征提取的图像语义分割网络，称为DResnet。首先提出一种双层残差网络，对训练 集各目标的细节进行特征提取，提高网络对部分细节目标的感知能力；其次在Lyel层开始跳跃特征融合，并 持续以2倍反卷积方法进行上采样操作，融合底层特征与高层特征，降低部分细节信息丢失对分割精度的影 响：最后使用网络分支训练法，先训练图像上各目标的大致轮廓特征，在此基础上再训练各目标的细节特征。 结果表明：该网络的平均交并比较全卷积网络相比，在CamVid上由49.72%提升至59.44%，在Cityscapes上由 44.35%提高到47.77%，该网络得到准确率更高、分割物体边缘更加完整的图像分割结果。 关键词：双层残差网络：细节特征提取：跳跃特征融合；上采样：网络分支训练法；图像语义分割：CamVid数据 集；Cityscapes数据集 中图分类号：TP183文献标志码：A文章编号：1673-4785(2022)04-0780-08 中文引用格式：谭睿俊，赵志诚，谢新林.双层残差语义分割网络及交通场景应用J.智能系统学报，2022,17(4)：780-787. 英文引用格式：TAN Ruijun,,ZHAO Zhicheng,XIE Xinlin.Double-residual semantic segmentation network and traffic scenic ap- plication[J].CAAI transactions on intelligent systems,2022,17(4):780-787. Double-residual semantic segmentation network and traffic scenic application TAN Ruijun,ZHAO Zhicheng,XIE Xinlin (School of Electronic Information Engineering,Taiyuan University of Science and Technology,Taiyuan 030024,China) Abstract:An image semantic segmentation network based on the double-residual network,named DRsenet,is pro- posed to address the depth problem of feature extraction network in semantic image segmentation and resolution reduc- tion of feature maps by a down-sampling pooling layer.Firstly,a double-residual network is proposed to extract the de- tailed features of each target in the training set and improve the perception ability of the network for some detailed tar- gets.Secondly,the jump feature fusion starts from Layer 1,and the up-sampling operation is continued by the 2xdecon- volution method to fuse the low-level and the high-level features to reduce the impact of partial detail information loss on the segmentation accuracy.Finally,the network branch training method is adopted to train the outline features of each target on the image,followed by training the detailed features of each target.The results indicate that the network's PMIou improves from 49.72%to 59.44%on CamVid and from 44.35%to 47.77%on Cityscapes when compared to the full convolution network.The network can produce image segmentation results with higher accuracy and more com- plete edge segmentation. Keywords:double-residual network;detail feature extraction;jump feature fusion;up sampling;network branch train- ing;semantic image segmentation;CamVid data set,Cityscapes data set 近年来，随着深度学习的不断发展，计算机视觉领域取得显著的成就。越来越多的交通场景需 要精确且高效的分割技术，实现交通场景的图像 收稿日期：2021-06-13.网络出版日期：2022-06-08 基金项目：山西省自然科学基金青年基金项目(201901D211304). 语义分割成为广大人员研究的问题之一。交通场 通信作者：赵志诚.E-mai:zhzhich@126.com. 景的图像语义分割，是对交通场景图像底层的像DOI: 10.11992/tis.202106020 网络出版地址: https://kns.cnki.net/kcms/detail/23.1538.TP.20220608.1352.002.html 双层残差语义分割网络及交通场景应用 谭睿俊，赵志诚，谢新林 （太原科技大学 电子信息工程学院，山西 太原 030024） 摘 要：针对图像语义分割过程中特征提取网络的深度问题以及下采样池化层降低特征图分辨率等问题，提出 了一种基于双层残差网络特征提取的图像语义分割网络，称为 DResnet。首先提出一种双层残差网络，对训练 集各目标的细节进行特征提取，提高网络对部分细节目标的感知能力；其次在 Layer1 层开始跳跃特征融合，并 持续以 2 倍反卷积方法进行上采样操作，融合底层特征与高层特征，降低部分细节信息丢失对分割精度的影 响；最后使用网络分支训练法，先训练图像上各目标的大致轮廓特征，在此基础上再训练各目标的细节特征。 结果表明：该网络的平均交并比较全卷积网络相比，在 CamVid 上由 49.72% 提升至 59.44%，在 Cityscapes 上由 44.35% 提高到 47.77%，该网络得到准确率更高、分割物体边缘更加完整的图像分割结果。 关键词：双层残差网络；细节特征提取；跳跃特征融合；上采样；网络分支训练法；图像语义分割；CamVid 数据 集；Cityscapes 数据集 中图分类号：TP183 文献标志码：A 文章编号：1673−4785(2022)04−0780−08 中文引用格式：谭睿俊, 赵志诚, 谢新林. 双层残差语义分割网络及交通场景应用 [J]. 智能系统学报, 2022, 17(4): 780–787. 英文引用格式：TAN Ruijun, ZHAO Zhicheng, XIE Xinlin. Double-residual semantic segmentation network and traffic scenic ap￾plication[J]. CAAI transactions on intelligent systems, 2022, 17(4): 780–787. Double-residual semantic segmentation network and traffic scenic application TAN Ruijun，ZHAO Zhicheng，XIE Xinlin (School of Electronic Information Engineering, Taiyuan University of Science and Technology, Taiyuan 030024, China) Abstract: An image semantic segmentation network based on the double-residual network, named DRsenet, is pro￾posed to address the depth problem of feature extraction network in semantic image segmentation and resolution reduc￾tion of feature maps by a down-sampling pooling layer. Firstly, a double-residual network is proposed to extract the de￾tailed features of each target in the training set and improve the perception ability of the network for some detailed tar￾gets. Secondly, the jump feature fusion starts from Layer 1, and the up-sampling operation is continued by the 2×decon￾volution method to fuse the low-level and the high-level features to reduce the impact of partial detail information loss on the segmentation accuracy. Finally, the network branch training method is adopted to train the outline features of each target on the image, followed by training the detailed features of each target. The results indicate that the network’s PMIOU improves from 49.72% to 59.44% on CamVid and from 44.35% to 47.77% on Cityscapes when compared to the full convolution network. The network can produce image segmentation results with higher accuracy and more com￾plete edge segmentation. Keywords: double-residual network; detail feature extraction; jump feature fusion; up sampling; network branch train￾ing; semantic image segmentation; CamVid data set; Cityscapes data set 近年来，随着深度学习的不断发展，计算机视 觉领域取得显著的成就。越来越多的交通场景需 要精确且高效的分割技术，实现交通场景的图像 语义分割成为广大人员研究的问题之一。交通场 景的图像语义分割，是对交通场景图像底层的像 收稿日期：2021−06−13. 网络出版日期：2022−06−08. 基金项目：山西省自然科学基金青年基金项目（201901D211304）. 通信作者：赵志诚. E-mai：zhzhich@126.com. 第 17 卷第 4 期 智 能 系 统 学 报 Vol.17 No.4 2022 年 7 月 CAAI Transactions on Intelligent Systems Jul. 2022