Research On Remote Sensing Road Image Semantic Segmentation Based On DeepLabV3+

With the rapid development of remote sensing technology,extracting information from remote sensing images has become one of the important ways for people to obtain information.Road information,as the most critical geographic information in remote sensing images,has been widely used in urban and rural road planning,automatic map updating,and smart transportation.Therefore,accurately and efficiently segmenting roads from remote sensing images has significant implications.In recent years,some deep learning-based image semantic segmentation models(such as Unet,Unet++,and Deeplab)have been proposed and have shown good performance in some scenarios.However,these models still face some accuracy and performance issues when dealing with complex remote sensing images.To address some challenges in semantic segmentation of remote sensing road images,this thesis presents the following research work:(1)To address the problem of the complex structure and large floating-point calculations of the DeepLabV3+ semantic segmentation model,a lightweight network,Mobile Net V2,is proposed as the feature extraction network of the semantic segmentation model.This approach can effectively reduce the model’s parameter and FLOPs,improve the segmentation efficiency,and introduce Mix Conv in Mobile Net V2 to extract more abundant feature information through different sizes of convolution kernels,which is beneficial to improve the model’s segmentation accuracy.(2)To address the inconsistency problem between the trained and predicted models caused by Dropout in DeepLabV3+,the R-Drop regularization method is proposed to further constrain the model.By constraining the output results of the same data sample through the model twice,the difference between the trained and predicted models is reduced,which is beneficial to improve the model’s segmentation accuracy and generalization performance.(3)The proposed optimization method is tested using the Deep Globe dataset,and the effectiveness is proven.The optimized model has better segmentation accuracy,with an MIo U improvement of 3.60%.In addition,the generalization experiment is conducted on the CHN6-CUG and Massachusetts datasets,and the results show that the MIo U is improved by 3.13% and 2.63%,respectively,which fully demonstrates the good generalization performance of the proposed optimization method.