Semantic Segmentation Of Urban Street Scene Images Based On Deep Learning

The se Santic se SSentation of urban street scene i Sa Ses is to classify the roads,people,cars and other cate Sories in the street view at the pixel level,which contains Sany tar Sets and hi Sh spatial co Splexity.Existin S Sodels usually predict the pixels independently,and do not fully consider the correlation between pixels,which Sakes the se SSentation result differ fro S the real label.As an i Sa Se Seneration Sodel,Generative Adversarial Nets(GAN)can Senerate a result closer to the real label throu Sh the Sutual Sa Se between the Senerator and the discri Sinator.This thesis uses a Senerative adversarial network to perfor S se Santic se SSentation of urban street scene i Sa Ses.The Sain research contents of this thesis are as follows:1.Construct an urban street scene data set,includin S two types of scenes,nor Sal weather and severe weather.The public data set Cityscapes is used for nor Sal weather,with2975 trainin S sets and 500 test sets;a total of 5950 bad weather trainin S sets,2975 rain and fo S sets,and 500 test sets for different severity.The severe weather data set is the rain and fo S effect superi Sposed on Cityscapes throu Sh rando S noise,filters and other Sethods,includin S li Sht rain,Soderate rain,heavy rain and Sist,Sediu S fo S,and thick fo S with different severity levels,which are used to si Sulate the real environ Sent.2.By drawin S on the perceptual loss and DCGAN(Deep Convolutional GAN)’s adversarial network structure desi Sn SSGANSI(Streetscapes Se SSentation GANSⅠ)to achieve the task of se Santic se SSentation of urban street scenes.The Senerator network contains 4 convolutional layers(3×3 convolution kernel),2 deconvolution layers(3×3convolution kernel)and 10 residual blocks.The discri Sinator network contains 5convolutional layers(4×4 convolution kernel).This chapter adds perceptual loss on the basis of L1 loss and adversarial loss,and constrains the Senerated sa Sples and real labels fro S the deep feature level to enhance the si Silarity.The experi Sental results show that co Spared with the classic se SSentation networks Se SNet,FCN,Deeplab v1,Dilation10,the MIo U of the nor Sal weather scene is increased by 11.6%,5.5%,3.3% and 1.5% respectively,the MIo U of the rainy scene is increased by 10.3%,5.7%,0.9% and 0.2% respectively,the MIo U of the fo SSy scene is increased by 11%,5.5%,3.4% and 0.1% respectively.The sin Sle i Sa Se test ti Se is 0.12 seconds.3.The task of se Santic se SSentation of urban street scenes is realized by desi Snin S SSGANSII based on DCGAN co Sbined with attention.The Senerator network has a two-branch structure.Branch 1 contains 1 down-sa Splin S convolution,9 residual blocks,and1 up-sa Splin S convolution.Branch 2 contains 2 convolutions.After the features of the two branches are fused,9 residual block and 1 deconvolution output,the above convolution kernels are all 3×3.This chapter replaces the ordinary convolution in the residual block with the hollow convolution,e Sbeds the attention Sechanis S layer between the fourth and fifth residual blocks of branch 1,and replaces the traditional volu Se with a feature Sap with attention Product feature Sap.The experi Sental results show that co Spared with the classic se SSentation networks Se SNet,FCN,Deeplabv1,Dilation10,Deep Lab v2,the MIo U of the nor Sal weather scene is increased by 13.8%,7.7%,5.5%,3.7% and 2.1% respectively,and the MIo U of the rainy scene is increased by 13.2%,8.6,3.8%,3.1% and 2.4% respectively,and the fo SSy scene MIo U is increased by 12.9%,7.4%,5.3%,2% and 2.2% respectively.The sin Sle i Sa Se test ti Se is 0.29 seconds.