| Image semantic segmentation is a very hot research direction,it is very important for scene understanding.In recent years,the semantic segmentation of streets based on deep learning has made great progress,especially with the rapid development of cameras,which produce a lot of street view images.If the computer can carry out accurate semantic segmentation of the image,it can lay a solid foundation for urban applications such as automatic driving and intelligent services.Therefore,this paper uses the school surveillance camera to acquire images,and conducts semantic segmentation research on these street view pictures.Firstly,the purpose and meaning of street semantic segmentation are introduced in detail,and the difficulty of it is explained.Secondly,semantic segmentation technology is divided into the segmentation before the emergence of deep learning,that is,traditional segmentation.Traditional segmentation is combined with the segmentation based on deep learning,and the segmentation based on deep learning is mainly introduced,and various aspects of knowledge related to deep convolutional neural networks are introduced.Several mainstream semantic segmentation models are also introduced.Google launches the Deeplab series and concludes with Deeplab V3+[1].This paper proposes a network model based on Deeplab V3+.Deeplab V3+ is a network with two parts: an encoder and a decoder.Among them,the encoder is mainly formed by the backbone feature extraction network and ASPP module together,and in the encoder,the pooling operation is replaced by the empty convolution with step size of 2,and the depth separable convolution is also used extensively to achieve the operation of reducing operation.Compared to encoders,decoders are relatively simple,using some simple ordinary convolution.After the decoder part passes through the ASPP module,it first goes through a common convolution and then through a quadruple upsampling module to fuse with the feature graph extracted from the backbone extraction network,and then through a common convolution and quadruple upsampling module to restore the original resolution.In this article,Deeplab V3+ network will be improved by replacing the backbone network and the decoder part.When training the network,it is found that the number of Deeplab V3+ parameters is too large,which leads to too long training time.In order to solve this problem,it is decided to introduce the popular idea of transfer learning and add pre-training weight to the network training process,which can not only speed up the training time,but also enable the model to obtain more features.In addition,the original Deeplab V3+network and other mainstream networks are compared,and the experimental results show that both the convergence speed and accuracy are significantly improved.Finally,the experiment is divided into three parts:(1)Mobile Net-V2[2] and Xception[3] were respectively used in the trunk extraction network.The test accuracy of Mobile Net-V2 is much higher than that of Xception and the number of parameters of Mobile Net-V2 is much smaller than that of Xception.(2)Adjustment and replacement of decoder,the new decoder is compared with the decoder of the original network,and the results show that the new decoder has better effect than the original decoder.(3)Transfer training was carried out on the network models respectively.The experimental results show that the network models after transfer training can achieve higher MIoU value. |
