Research Of Road Extraction Of Remote Sensing Image Based On Generative Adversarial Network

With the increase of infrastructure construction and the improvement of road network system in China,road extraction through remote sensing image has become an important research content and has great application value.However,the existing road extraction methods still have some problems,such as low extraction accuracy,low automation and needing a large number of marking data.Based on the above problems,this paper regards the road extraction of remote sensing image as an image conversion,and proposes a road extraction method of remote sensing image based on the generated adversarial network,aiming at solving the problem of too much data labeling and data preprocessing and insufficient automation with a high level of accuracy.The work of this paper is summarized as follows:First,aiming at the task of map style image generation,a method of generating adversarial network based on depth convolution is proposed,and the experiment of map style generation is carried out.This experiment is a prior experiment of road extraction from remote sensing image,which is used to verify the feasibility of map style image generation.The experimental results show that the method proposed in this chapter can effectively generate map style images,which provides a strong support for the feasibility of later research.Secondly,a road extraction method of remote sensing image based on pix2 pix generation adversarial network is proposed for the road extraction task of paired remote sensing image.The road extraction task is regarded as a special image conversion task.The remote sensing image is transformed into the corresponding map style image by generate adversarial network,and then the road network can be extracted by simple RGB value judgment and morphological processing to reduce the difficulty of road extraction.The experimental results show that the accuracy is94.26%,recall is 81.08%,precision is 84.62%,F1 score is 84.05%,which is a practical and effective road extraction method.Thirdly,a road extraction method based on cycle generation adversarial network is proposed for the road extraction task of unpaired remote sensing image.There are two groups of generate adversarial networks in the model of cycle generate adversarial network,which respectively complete the transformation from data domain x to data domain y and from data domain y to data domain X.the two groups of generate adversarial networks form a cycle,using cyclic loss as loss function,ensuring cyclic consistency,thus solving the training problem of unpaired data.The experimental results show that the accuracy rate is 93.81%,the recall rate is 80.88%,the accuracy is 83.31%,F1 score is 85.38%,the overall road appearance effect is good,and the problem of road extraction from unpaired training data is solved when the evaluation index basically does not decline.