Study On Improvement And Lightweight Of Road Extraction Model In Remote Sensing Image

Deep learning,as a popular research content at present,has a wide range of applications in multidisciplinary intersection.Due to the rapid advancement of remote sensing technology,it is currently a hot topic in remote sensing to determine how to properly and quickly extract information such as features from remote sensing photos.As an important part of transportation,road is an important feature marker in remote sensing images,and has great applications in various fields of society,including urban planning,traffic management,military applications,and so on.Thus,it is crucial for study that road data be reliably and promptly extracted from remote sensing images.This thesis focuses on the task of road extraction from remote sensing images,based on convolutional neural network,and conducts systematic research and exploration from three aspects of improving road extraction accuracy,lightweight model and automatic extraction,and the mainly works and innovations are listed as follows:(1)In this study,we develop a road extraction network SA-Link Net based on D-Link Net coupled with an attention mechanism and chain inverted residual block.The attention mechanism operates on the network’s deep semantic information,which can pick the features that are most helpful for the task of road segmentation and suppress the irrelevant features to prevent task interference.The chain inverted residual block acts the network’s shallow spatial information,completes the fusion of shallow and deep context information,and improves the model’s expression capability,allowing it to more accurately differentiate between the road and the background.The network suggested in this research can have some improvement compared to the original model after being tested on the Deep Globe road extraction dataset,with F1-score and Io U score improving by 0.54% and0.88%,respectively,and some improvement in road disconnection and feature marker identification.(2)Two lightweight network models are suggested in order to address the issue that the semantic segmentation network is difficult to implement in edge computing devices due to its deep number of layers,numerous parameters,and amount of calculation.Inspired by Ghost Net,this thesis proposes lightweight components Ghost Conv Trans Module and Ghost Dilated Module.Two lightweight networks,G-Link Net and TG-Link Net,based on the D-Link Net network,were proposed by using lightweight components(Ghost Module,Ghost Conv Trans Module,and Ghost Dilated Module)to optimize the network’s parameters.Among them,both networks’ parameter counts are decreased,and TG-Link Net is the end result of additional G-Link Net lightweight.TG-Link Net combines Res Net34 with Ghost Module for parameter optimization in the encoder part,and proposes Ghost-Res Net34 encoder network.Compared with the original model,the network parameters of G-Link Net and TG-Link Net are reduced by 23.2% and 85.7%,respectively.Tested on the Deep Globe road extraction dataset,the F1-score and Io U score of G-Link Net are 0.79% and 1.14% higher than that of the original model.The comprehensive performance of TG-Link Net is better than that of the classical network U-Net,but it has fewer parameters and model file size,and better model performance.(3)In order to automate the extraction of remote sensing images,this research designs and develops a road extraction system for remote sensing images with a visual user interface for easy operation by users.Users can choose the appropriate network models for road extraction based on their requirements from the system’s collection of models,which also includes comparison models.Users can compare the models’ overall performance and visually compare the various extraction results among them.