Road Extraction From High Resolution Remote Sensing Images Based On Generative Adversarial Network

The high-resolution remote sensing images contain rich information,which lays a good foundation for its subsequent interpretation.Road extraction is an important research domain in high-resolution remote sensing image processing,and plays an irreplaceable role in many fields and scenes.However,the traditional road extraction method can't accurately extract discriminative information in the image,which leads to the inaccurate road extraction result of the traditional method and the low degree of automation and intelligence.However,the mainstream method based on deep learning has improved the extraction result,but it can't deal with the small sample problem in remote sensing image processing which requires a lot of data augmentation and pre-processing and post-processing operations that increases computing and storage resources consumption.Aiming at the problems above,this paper proposes a road extraction method based on semantic embedding Generative Adversarial Network(GAN),which improves the performance of road extraction under small sample conditions.A weak supervised multi-scale semantic embedding GAN is proposed to deal with the multi-level road extraction in high-resolution remote sensing images which improves the road extraction performance with large difference in road width,and further alleviates the limitation of the model effect due to small sample problem.A method to remove the occlusion on roads in remote sensing images based on GAN is proposed to avoid the negative impact on the road extraction result due to the occlusion problem that often occurs in remote sensing images,which enhances the robustness of the model in practical applications.The main content of this paper is summarized as follows:1.We propose a road extraction method based on semantic embedding GAN.This method regards the task of road extraction as a special task of image translation,which uses the generative network to generate road class maps from remote sensing images directly,and we add semantic embedding loss to original GAN loss to further constrain the quality of the generated image in detail.This method is a complete end-to-end road extraction framework,which can achieve excellent performance under the condition of small samples of remote sensing images.2.We propose a multi-level road extraction method based on weakly supervised multi-scale semantic embedding GAN.By using multi-channel generation network model,the multi-scale roads in remote sensing images are accurately extracted,and the number of parameters is reduced.Using progressive growing GAN to generate more high-resolution remote sensing image samples combined with the weak supervised strategy provides more labeled samples for deep network training,which alleviates the performance degradation of the model caused by small sample problems.3.We propose a method to remove occlusion on roads in remote sensing images based on GAN.This method abandons the idea of traditional remote sensing image to cloud,and uses the generated network to generate information directly of occlusion area,which can repair of missing information without prior knowledge of occlusion position.