无线电工程2024，Vol.54Issue(8)：1928-1935,8.DOI:10.3969/j.issn.1003-3106.2024.08.011
一种轻量化多尺度遥感图像分割方法
A Lightweight Multi-scale Remote Sensing Image Segmentation Method
摘要
Abstract
Semantic segmentation of remote sensing images plays an important role in urban planning and development.How to perform automatic,fast and effective semantic segmentation for highly complex and multi-class remote sensing images has become the key of research.However,the existing segmentation methods based on deep learning have the problems of complex model and high computational cost.An end-to-end lightweight Multi-Scale Feature Extraction and Segmentation Network(MSNET)is proposed to reduce the computational cost in the case of high accuracy.Firstly,the backbone is composed of a coding network based on lightweight network MobileNetV2 and a decoding network based on MSConv.MSConv is a new multi-scale convolutional module.In addition,a Feature Fusion Attention Module(MSAM)is proposed to effectively integrate the global information of attention mechanisms in channel and spatial dimensions.Secondly,a more lightweight Local Importance Pooling(LIP)is introduced to replace the common pooling operation,and the Atrous Spatial Pyramid Pooling(ASPP)module is added to further extract rich features.Finally,the comparison evaluation on the public dataset WHDLD showed that F1 reaches 83.12％and the inference time is only 0.007 4 s.关键词
遥感图像/图像分割/轻量化/多尺度Key words
remote sensing image/image segmentation/lightweight/multi-scale分类
信息技术与安全科学引用本文复制引用
雷帮军,余楷,吴正平..一种轻量化多尺度遥感图像分割方法[J].无线电工程,2024,54(8):1928-1935,8.基金项目
国家自然科学基金(61871258) （61871258）
水电工程智能视觉监测湖北省重点实验室建设(2019ZYYD007) （2019ZYYD007）
湖北省重点实验室开放基金(2018SDSJ05)National Natural Science Foundation of China(61871258) （2018SDSJ05）
Construction of Hubei Key Laboratory for Intelligent Visual Monitoring of Hydropower Engineering(2019ZYYD007) （2019ZYYD007）
Open Fund of Hubei Key Laboratory(2018SDSJ05) （2018SDSJ05）
