Road Scene Recognition Technology Based On U-Net Netwwork

The development of driverless technology is in full swing,and semantic segmentation is an important topic in driverless technology.Image semantic segmentation methods are divided into two stages: traditional methods and methods based on deep learning.The traditional method is relatively simple,subject to accuracy requirements,and the effect is not ideal.But later,deep learning was applied to semantic segmentation,especially the emergence of full convolutional networks,which enabled semantic segmentation to be processed at the pixel level,which greatly improved the accuracy of segmentation.In this paper,we mainly identify and segment the road edges to improve the edge accuracy.At the same time,three different weather conditions of rainy,sunny and evening are trained at the same time,so that the improved model can cope with various weather conditions.Based on the systematic study of various classic network models in recent years,this paper improves the original U-Net network model.Organize and label the data set,and divide the data set into three different states: sunny,rainy,and night for training at the same time,so that the model can cope with different weather conditions.At the same time,it combines residual calculation and splicing algorithms,and reuses features by introducing residual calculation modules to improve convergence speed and segmentation accuracy and prevent over-fitting,and the splicing algorithm is used to reuse the features learned in the shallow layer in the deep transfer process to reduce redundant features and greatly reduce training parameters to reduce the difficulty of training,while making full use of feature values.The output after the residual and splicing processing is used to replace the output of the encoder in the original U-Net network to ensure that the encoder and decoder in the overall structure correspond.Finally,semantic segmentation is performed on six network models,and the simulation results show that the segmentation accuracy is greatly improved.