Font Size: a A A

Research On Semantic Segmentation Method Of Remote Sensing Image Based On Deep Learning

Posted on:2022-10-19Degree:MasterType:Thesis
Country:ChinaCandidate:X XuFull Text:PDF
GTID:2512306530980109Subject:Electronics and Communications Engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of remote sensing technology,high-resolution remote sensing image has been widely used in environmental monitoring,disaster assessment and crop yield assessment.Semantic segmentation is one of the key tasks in computer vision,it can classify each pixel in the image.Because high-resolution remote sensing images have the characteristics of complex backgrounds and large differences in scales of different targets,traditional semantic segmentation methods are not ideal for remote sensing image processing.Deep learning technology has made great progress in recent years,and it is of practical significance to apply it to the analysis of remote sensing images.This paper studies the application of semantic segmentation method based on deep learning in high-resolution remote sensing images.The specific research contents are as follows:(1)According to the characteristics of high-resolution remote sensing images with complex backgrounds and large differences in scales of different targets,a highprecision multi-feature attention fusion network is proposed.The network is based on the encoding and decoding structure.In the encoding stage,the optimized deep residual network Res Net101 is used as the backbone network to extract features.Then the network captures the contextual information and channel-related information of the image through the pyramid pooling structure and the channel self-attention structure to enhance the feature representation ability of the output;in the decoding stage,the semantic information of the high-level features is used to guide the lowlevel features for effective fusion,and the cascading strategy is used to gradually restore and optimize the detailed information of the target.Comparative experiments were conducted on the public data sets Potsdam and Vaihingen.Compared with the high-precision semantic segmentation model PSPNet,it has achieved higher segmentation accuracy,and the pixel accuracy has been improved by 1.3% and 1%,respectively.(2)Aiming at the problems of complex computation and memory consumption in semantic segmentation models of high-resolution remote sensing images,a realtime two-branch lightweight network is proposed.The network contains two parallel branches,the detail branch and the semantic branch.In the detail branch,the detailed information of the target is captured through the shallow network of the wide channel;in the semantic branch,the context information of the image is captured through the deep network of the narrow channel.Then the network uses context-scale awareness to adaptively fuse the features of different branches.Finally,the intensifier and joint auxiliary loss function are used in the training process to speed up the convergence of the network.Comparative experiments were conducted on the public data sets Potsdam and Vaihingen.Compared with the real-time semantic segmentation model Bi Se Net,it achieves a similar segmentation accuracy.The amount of network parameters has been reduced by 76.1%,the amount of calculation has been reduced by 86.6%,and the inference speed has increased by 45.9%.
Keywords/Search Tags:High-resolution remote sensing images, Deep learning, Pyramid pooling, Channel self-attention, Real-time semantic segmentation
PDF Full Text Request
Related items