Research On Semantic Segmentation And Road Extraction Methods For Remote Sensing Images Based On Multi-scale Feature Extraction

In recent years,with the continuous development of air-space-space-sea integrated earth observation technology,the application of satellite remote sensing,big data,digital twin and real-world 3D technologies to build digital China and boost national high-quality development has become a very challenging task at present.Intelligent interpretation of remote sensing images has wide application in the fields of land survey,urban construction,intelligent agriculture,intelligent transportation and so on.As a key step of remote sensing intelligent interpretation,semantic segmentation of remote sensing images has been given much attention by researchers both domestically and abroad.With the advent of deep learning,convolutional neural networks have become a new research paradigm for semantic segmentation of remote sensing images.High-resolution remote sensing images,unlike natural images,are characterized by rich texture and detail information,multiple targets and multi-scale phenomena.Therefore,the focus of research is on how to accurately segment these targets from these images.In this paper,we mainly study the multi-scale feature representation and construct a semantic segmentation network to optimize the performance of semantic segmentation and road extraction tasks of high-resolution remote sensing images.The details are as follows:(1)This paper proposes a network based on multi-scale attention and non-local filters to address the issue of considerable feature target size variation in remote sensing images.This network extracts fine-grained multiscale features through a multiscale attention module,and utilizes a nonlocal filter to extract global contextual information.Experiments conducted on public datasets ISPRS Vaihingen and ISPRS Potsdam have demonstrated that this method is capable of significantly improving the accuracy of semantic segmentation in high-resolution remote sensing images.(2)In order to accurately extract roads with different scale variations in remote sensing scenes,this paper proposes a multi-scale gated full fusion module and constructs a road extraction network based on this module.By using a set of null convolutions with different null rates,the multiscale gated full fusion module enables the network to extract multiscale features at different scales,and by using nonlocal pooling,the network is supplemented with global information.In addition,the use of gated full fusion unit enables the network to effectively fuse features at different scales.Experiments on the publicly available datasets CHN6-CUG and DeepGlobe Road show that the proposed network can achieve good road extraction accuracy.The research in this paper improves the accuracy of semantic segmentation task of remote sensing images by different multi-scale feature extraction methods,and constructs a road extraction network for specific application tasks,which is expected to be applied to smart cities and smart transportation.