Research On Semantic Segmentation Technology Of Urban Road Based On Convolutional Neural Network

Image semantic segmentation technology labels different kinds of objects at pixel level,which provides abundant visual information for correct recognition and classification of objects.The segmented image is easily affected by illumination,occlusion,uneven sample distribution and other factors,which further affects the segmentation effect of the model.In the field of automatic driving,in order to ensure the safe driving of vehicles,automatic driving system is needed to provide real-time and accurate driving area around the road for vehicles to make correct operation.Due to the complex urban road environment,the existing network model has low segmentation accuracy,misjudgment,missed judgment and other phenomena.Moreover,the network model is complex,with a large amount of computation and time-consuming image segmentation,which cannot meet the actual needs of autonomous vehicles on the mobile side.In order to improve its segmentation accuracy and speed,two improved network models are studied and proposed.The specific contents and innovations are as follows.1.In the semantic segmentation network model Deeplabv3+,the segmentation images are affected by illumination,overlap,occlusion and other factors,which makes the segmentation results of the model easy to miss small objects,easy to misjudge different types of objects with similar shapes,and fuzzy segmentation boundaries.Therefore,attention mechanism is introduced,and a large amount of high-level context information and low-level semantic information can be captured by using position attention module and channel attention module,and the addition of attention module will not increase the parameters and computation of network model.A semantic segmentation method of road scene integrating attention mechanism is proposed.A set of parallel channel attention modules and position attention modules are introduced at its encoding and decoding ends to capture and output more refined results.Meanwhile,BN operation is added to normalize the data,which accelerates the convergence speed of the model and improves the segmentation accuracy of the model.2.Aiming at the problems that the current image semantic segmentation model has complex network structure,large amount of computation,and it is difficult to realize its use on the mobile terminal,a lightweight real-time semantic segmentation network model GINet for urban road scenes is constructed.GINet is based on the coding-decoding structure,adopts Mobile Netv2 structure at the coding end,introduces inverse residual unit,constructs a lightweight network,and introduces expansion convolution in inverse residual unit to increase the receptive field of the network and shorten the training time of the model.By introducing the global average pooling layer,the whole connection layer in the network is replaced,and a large amount of global information is captured to improve the segmentation accuracy of the model.At the decoding end,the image features are recovered by bilinear interpolation upsampling operation and the refined segmentation results are output.The effectiveness of the design method in this paper is verified by the experiments of the above two network models on the large and open road scene dataset Cam Vid dataset and Cityscapes dataset.