Research On Road Scene Semantic Segmentation Technology Based On Deep Learning

In the field of computer vision,semantic segmentation enables the classification of images at the pixel level.It is able to take a picture or video and divide the image into multiple blocks according to the similarities and differences of the categories.The development of semantic image segmentation techniques has inspired many scholars to research.At the same time semantic segmentation technology has a wide range of applications in many fields,such as medical imaging,remote sensing images,autonomous driving and so on.For autonomous driving,the semantic segmentation of road scenes has become particularly important.However,the output effect of existing semantic segmentation techniques for road scenes has problems such as low prediction accuracy and slow training speed.Therefore,this paper analyses the characteristics of the previous network structure and proposes an improved technique for the application in road scenes.(1)To address the problem of low accuracy of the traditional U-Net model for semantic segmentation of road scenes,this paper proposes a Du-Net network based on the fusion of DeepLabV3+ and U-Net for semantic segmentation of road scenes.Du-Net is composed of a DeepLabV3+ network structure in the first stage and a U-Net based on the attention module of ECA channels in the second stage.Net is constructed.The output of the segmented image input to DeepLabV3+ is used as input to the second stage of the U-Net based on the ECA channel attention module,which allows the road edge images to be more granular.The application to the road scenes results in a significant improvement in the accuracy of the road edge segmentation of the images.(2)In order to reduce the influence of complex background on the semantic segmentation of road images,The Efficient Channel Attention is embedded between the three convolution and pooling operations in the coding part of the second stage U-Net,which better improves the adjustment ability of the feature channels in the coding stage and improves the segmentation accuracy to a certain extent.(3)The segmentation speed of semantic segmentation of road scenes is a consideration that cannot be ignored in autonomous driving technology.In this paper,the lightweight MobileNet V2 is used in the backbone network of DeepLabV3+ in the first stage,and the four convolution and pooling operations used in the coding stage are changed to three times in the second stage of the U-Net network structure,and the four upsampling and convolution operations used in the decoding part are changed to three times in order to improve the speed of semantic segmentation,which provides the possibility for the Du-Net semantic segmentation technology to be used in The possibility of deploying Du-Net semantic segmentation technology for autonomous driving is provided.Finally,the dataset was annotated using the Labelme annotation tool and the results were compared experimentally between the road semantic segmentation based on Du-Net and other network structures.