Research And Implementation Of Improving Semantic Segmentation Effect Of Traffic Scene Image Based On Attention Mechanism

With the deepening of the research on automatic driving technology,the in-depth understanding of traffic scene becomes more and more important.The semantic segmentation of traffic scene image plays an important role in the field of automatic driving.Only by effectively improving the accuracy of semantic segmentation,can the automatic driving vehicle navigate safely in the complex daily traffic scene.At present,many semantic segmentation methods have low segmentation ability for small objects and the boundary of objects,and automatic driving requires high segmentation accuracy for small objects.Therefore,the research on the accurate segmentation technology of small objects and the boundary edge of objects will have a very strong role in the field of automatic driving.This paper aims to study the improvement of traffic scene image segmentation effect after embedding attention module in semantic segmentation neural network.In order to improve the accuracy of semantic segmentation,the main research contents of this paper are as follows.1.In order to improve the accuracy of network segmentation of small objects in traffic scene images,the spatial attention module is added based on Deep Labv3+semantic segmentation network to enhance the feature expression ability of spatial feature map;the structure of ASPP module is optimized to expand the range of sensing field and effectively improve the segmentation performance of small objects in traffic scene images;at the same time,the parameter selection of Focal Loss function can effectively alleviate the impact of unbalanced data and increased sensitivity due to the addition of attention mechanism module.Finally,the m Io U of the network in cityscapes test set is 78.56%.Compared with Deep Labv3+,the network segmentation performance is improved while the average prediction time is basically unchanged.2.In order to improve the accuracy of semantic segmentation at the boundary of objects in traffic scenes,on the basis of the Deep Labv3+ semantic segmentation algorithm,a channel attention module is added to the design.In addition to the Focal Loss function,a Multi-ASPP module is also designed.This module divides the feature map into average pooling channels and maximum pooling channels.The ability of Deep Labv3+ to express channel features is enhanced,and the channel features are further enriched which improves the accuracy of semantic segmentation of the boundary edges of objects in traffic scenes.Finally,the addition of the Multi-ASPP module has increased the network m Io U from 77.17% of Deep Labv3+ to 78.82%,and the average prediction time has been reduced from 0.150 s of Deep Labv3+ to 0.146 s,which improves the accuracy of performance and improves network efficiency.3.In order to further improve the accuracy of network segmentation for small objects and the boundary edge of objects in the traffic scene image,the spatial attention module and channel attention module are combined in parallel,and added to Deep Labv3+,so that the model effectively integrates the spatial features and channel features.The non-local attention module is introduced into the backbone network to form a non-local network,which enhances the ability of the backbone network to extract features,and further improves the accuracy of the network for the segmentation of small objects and the boundary edge of objects in the traffic scene.The m Io U of the final network is 79.42%.