Research On High Precision Real-time Semantic Segmentation Algorithm For Road Scenes Based On Deep Convolutional Neural Network

Semantic segmentation algorithm for road scene image is a common technology to realize automatic driving environment perception.In recent years,with the development of Deep Learning,the segmentation accuracy of image semantic segmentation algorithm based on Deep Convolutional Neural Network is higher and higher,but the algorithm structure is more and more complex.At present,it is difficult to meet the real-time requirements of automatic driving by running such a complex segmentation network.To solve the above problems,two high-precision real-time semantic segmentation algorithms for road scenes based on Deep Convolutional Neural Network are proposed and they are BFBNet and BANet.To verify the correctness of the two algorithms,ablation experiment and comparison experiment were carried out.Experimental results show that the two algorithms maintain high segmentation accuracy when the speed meets the requirement of real-time,and BANet is slightly better than BFBNet in segmentation accuracy and segmentation speed.The main contents of this paper are as follows:In this paper,the first algorithm is Bilateral Real-time Semantic Regmentation Network Based on Bilateral Fusion(BFBNet).Firstly,to improve the segmentation speed of the network,BFBNet adopts a dual-branch network structure.One branch is semantic branch,which aims to extract high-level and low-resolution semantic feature information of images.The other branch is spatial detail branch,which aims to extract low-level and high-resolution spatial detail feature information of images.Secondly,to improve the ability of extracting semantic features and spatial detail features,lightweight Multi-scale Information Fusion Module and Bilateral Unidirectional Fusion Module are added to semantic branch and spatial detail branch respectively.Finally,in order to fuse the output feature information extracted from the two branches,a bilateral bidirectional fusion module is designed to fuse the information of the two branches.On Cityscapes dataset,the proposed BFBNet network is validated.The results show that the MIo U value of the proposed algorithm can reach 73.75%,and the number of images transmitted per second can reach 77.93 FPS.The second algorithm proposed in this paper is Bilateral Real-time Semantic Regmentation Network Based on Attention Mechanism(BANet).First,like BFBNet,BANet has a dual-branch structure.Secondly,to make BANet network pay attention to both channel domain information and spatial domain information,to enhance the ability of semantic branch to extract semantic information,a hybrid domain attention module is added to semantic branch.At the same time,to make the spatial detail branch of BANet pay more attention to spatial information,the spatial attention mechanism module is added to the branch.Finally,the MIo U value of BANet network verified on Cityscapes dataset can reach 74.23%,and the image transmission frame number can reach 90.63 FPS.Figure 40,table 14,reference 41.