Research On Semantic Segmentation Of Autopilot Based On Fully Convolutional Network

The advances of technology have made autonomous driving gradually become a reality from people’s imagination.To achieve this goal,it depends on the analysis of road scenes,which requires semantic segmentation technology.The semantic segmentation technology is used to classify each pixel in the image to complete the analysis of the road scene.The early image segmentation methods have many defects,and later deep learning has opened up a new approach for it.In recent years,convolutional neural networks have been widely used in semantic segmentation tasks,with outstanding performance in the field of image segmentation.The proposal of fully convolutional neural network has a milestone significance in the development of semantic segmentation.The semantic segmentation technology based on the fully convolutional neural network is more flexible,smarter and superior in performance compared to the earlier image segmentation methods.Unlike the traditional convolutional neural network,the final output is a vector.The fully convolutional neural network can accept any size of input and get the same size of the output feature map.It is an end-to-end network architecture.Based on the fully convolutional neural network,this paper improves the semantic segmentation network by using methods such as skip connection,atrous convolution and multi-layer feature fusion.This article focuses on two core topics: to propose a simple new fully convolutional neural network;to improve the performance of semantic segmentation network.The research content of this article is as follows:Aiming at the problem of loss of feature information caused by the current convolutional neural network using pooling layer,a simple full convolutional network model is proposed.Unlike traditional convolutional neural networks,this model removes the pooling layer,and the entire model is mainly composed of convolutional layers.On the premise of retaining the spatial dimensionality reduction function,the convolutional layer with the same function is used to replace the pooling layer to achieve the purpose of retaining more feature information.Then,the feature fusion of different network layers is completed through jump connections,and the semanticinformation of the deep layer of the network is combined with the detailed information of the shallow layer of the network to obtain better prediction results.Experimental data shows that the prediction accuracy of the proposed network model is better than the basic fully convolutional neural network.Aiming at the problem that the feature resolution is too low due to multiple downsampling of traditional convolutional neural network to obtain a larger receptive field,a multi-layer feature fusion network model with atrous convolution is proposed.The network model keeps more feature information by reducing the downsampling operation layers so that the feature resolution will not be too low.At the same time,in order to make up for the problem of reducing the insufficient receptive field caused by downsampling,atrous convolution is used to expand the receptive field,increase the range of acquired feature information,and enrich the feature content.The deconvolution part uses a multi-layer fusion method to further merge the features of different network layers,increase the details of feature mapping,and improve the accuracy of network prediction.