Urban Scene Semantic Segmentation Based On Convolutional Neural Network

At present,the main scenes in the field of autonomous driving and drone distribution are urban scenes.In the field of autonomous driving,real-time analysis of the surrounding environment of the vehicle is required to ensure safe driving of the vehicle,as well as in the field of drone distribution.With the rapid development of the autonomous driving industry and the distribution of drones,the demand for segmentation of the surrounding environment has also increased day by day.Therefore,the research on segmentation of urban scenes has very important practical significance.At present,the semantic segmentation technique can be profiled in two steps: the first step is to extract image features by convolution operation,and the image information is extracted and understood;the second step is to restore the upsampling operation of the feature image.At present,most of the research on semantic segmentation focuses on the method of segmentation accuracy improvement,such as DeepLab series.There are not many researches on segmentation efficiency.However,in practical life applications,segmentation efficiency is also an important part.This paper proposes a real-time semantic segmentation network model based on densely connected networks to complete the study of urban scenes.In the network model,the dense connection block is used repeatedly to extract the feature information,and the latest research results such as asymmetric convolution and expansion convolution are used in the dense connection block,so that the model can efficiently and accurately complete the segmentation task of the urban scene.The Cityscapes dataset is used to complete the training and testing of the model on the PyTorch framework.The parameters and structure of the model are selected through experiments,including growth rate,decoder structure,and pooling structure of the void space.The method of simple fusion of models and the method of phased training are also used to optimize the segmentation precision.Finally,the model and other related semantic segmentation network models are compared in terms of segmentation accuracy and running time.The comparison results show that this model is one of the best algorithms in terms of integration time and precision.