双层残差语义分割网络及交通场景应用  

Double-residual semantic segmentation network and traffic scenic application

作　　者：谭睿俊 赵志诚[1] 谢新林 TAN Ruijun;ZHAO Zhicheng;XIE Xinlin(School of Electronic Information Engineering,Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学电子信息工程学院,山西太原030024

出　　处：《智能系统学报》2022年第4期780-787,共8页CAAI Transactions on Intelligent Systems

基　　金：山西省自然科学基金青年基金项目(201901D211304).

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。An image semantic segmentation network based on the double-residual network,named DRsenet,is proposed to address the depth problem of feature extraction network in semantic image segmentation and resolution reduction of feature maps by a down-sampling pooling layer.Firstly,a double-residual network is proposed to extract the detailed features of each target in the training set and improve the perception ability of the network for some detailed targets.Secondly,the jump feature fusion starts from Layer 1,and the up-sampling operation is continued by the 2×deconvolution 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 complete edge segmentation.

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 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]