摘要
交通场景语义分割在自动驾驶中必不可少。为了解决目前的交通场景语义分割方法中由于池化、卷积等操作而造成的目标边界分割模糊、多类别目标及相似物体分割精度低等问题,提出一种带注意力机制的卷积神经网络分割方法。特征提取时,引入多样化的扩张卷积,以挖掘多尺度的语义信息。在信息解码后,添加通道及空间双注意力模块,可以在通道和空间两个维度层面进行注意力特征提取,让网络在学习过程中更侧重于重要信息。在Cityscapes数据集上的实验结果表明,该语义分割网络的平均交并比(MIoU)可达71.6%,超过了基网络为ResNet50的DeepLabv3+语义分割网络。所提方法能更加精细地分割出近似物体及多类别目标,对复杂交通场景图像的理解力更强。
Semantic segmentation of traffic scenes is an essential part of automatic driving.In the current traffic scene semantic segmentation methods,due to operations such as pooling and convolution,the segmentation of target boundaries is blurred,and the segmentation accuracy of multi-category targets and similar objects is low.For the issue above,a segmentation method based on convolutional neural network with attention mechanism was proposed.In feature extraction,diversified dilated convolutions were introduced to extract multi-scale semantic information.After the information was decoded,channel and spatial dual attention modules were added to extract attention features at the channel and spatial dimensions,so as to make the network focus more on the important information in the learning process.Experiments on the Cityscapes dataset show that the Mean Intersection over Union(MIoU)of the proposed semantic segmentation network can achieve 71.6%,which exceeds the DeepLabv3+semantic segmentation network whose base network is ResNet50.The proposed segmentation method can segment similar objects and multi-category targets more finely,and has a stronger understanding of complex traffic scene images.
作者
何淼楹
崔宇超
HE Miaoying;CUI Yuchao(College of Electronics and Information Engineering,Sichuan University,Chengdu Sichuan 610041,China)
出处
《计算机应用》
CSCD
北大核心
2021年第S01期25-30,共6页
journal of Computer Applications
基金
四川省重点研发计划项目(2019YFG0409)。
