Research On Remote Sensing Image Segmentation Along Highway Based On Generative Adversarial Network

As one of the important infrastructures driving regional economic development,highway plays an important role in the development of national economy.With the country’s vigorous promotion of transportation power,remote sensing technology is used for environmental monitoring along the highway.Semantic segmentation of remote sensing images can help staff better grasp the ground objects along the highway and help carry out dynamic monitoring and safety analysis of the environment along the highway.It plays an important role in earth observation,environmental monitoring and disaster early warning.In recent years,the image semantic segmentation method based on deep learning can obtain the required image semantic information efficiently and accurately,which has great practical value.However,due to the complex background of remote sensing image data sets along the highway,the existing semantic segmentation models have some phenomena,such as rough boundary contour,wrong segmentation,missing segmentation and so on;Moreover,such methods require a large number of sample annotations,which is a challenge for semantic segmentation tasks.Obviously,highway construction has the characteristics of wide coverage and long route.The roads in different regions and different local sections of the same highway have different ground features,such as cities,villages,market towns and other areas with large population and dense buildings,and some remote areas with few people and mainly trees,wasteland and Gobi,which have different requirements for the segmentation accuracy of ground features.Based on this,this research uses the generative adversarial network(GAN)to realize the semantic segmentation of fully supervised learning image and semi supervised learning image.The main work and innovation of this research include the following three aspects:(1)Semantic segmentation of remote sensing image based on improved pix 2 pix.The supervised learning model Pix2 Pix is used to realize semantic segmentation in the area where the segmentation accuracy of detailed features is high.In view of the loss of detail information caused by the down sampling operation of the segmentation network of the classical pix 2 pix model,which leads to the decline of segmentation accuracy,this research makes the following improvements.Firstly,at the end of the coding stage,connect the atrus spatial pyramid pooling(ASPP)module to capture the object and image context in multiple scales,and introduce spatial attention mechanisms(SAM)to enhance the edge detail information of the ground object,so as to improve the segmentation ability of the network model.(2)Semi supervised remote sensing image semantic segmentation based on improved cycle consistent generative adversarial networks(Cycle GAN).For remote areas with few people and single objects,semi supervised learning is used for image semantic segmentation.At first,Sam is introduced into the convolution layer of each feature extraction stage to enhance the feature extraction ability,then ASPP is connected to capture multi-scale information at the end of feature extraction,and semi supervised training method is used for training.Using less training data,the segmentation accuracy of the model is improved,and the time and labor cost of labeling are greatly reduced.(3)Design and implementation of semantic segmentation system along highway.The requirements of the system are analyzed from the aspects of business,functionality and performance.On this basis,the front-end and back-end adopt Lay UI framework and Spring MVC + Spring + Mybatis framework to design and realize the functions of data import,image segmentation,segmentation result statistics,segmentation result reporting,approval and approval status tracking.The semantic segmentation system along the highway has simple operation,stable performance and high practical value.The experimental results show that this method effectively improves the accuracy of semantic segmentation and alleviates the problems of wrong segmentation,missing segmentation and fuzzy edge segmentation.The division of environmental features along the highway is more accurate,and it is applied to the system to better improve the efficiency of environmental monitoring along the highway.