摘要
针对高分辨率遥感影像地物分割研究普遍存在的计算复杂、分割精度低、空洞大等缺陷,提出了引入双注意力机制的DeepLabV3+网络模型。该模型采用轻量化MobilenetV2作为主干网络,并充分考虑了网络参数设定、边缘提取优化和性能三个方面,从而获取密集的上下文信息。将所设计的模型应用于寻乌县高分辨率遥感影像数据集验证,结果表明,加入双注意力机制的DeepLabV3+网络模型对八种地物的分割均取得较好的分割精度,尤其是对园林、水体和道路的分割,分割精度高达92%、90%和96%。本研究为高分影像的地物分割及如何弥补基础的DeepLavV3+缺陷等问题提供科学参考。
Aiming at the common shortcomings of the research on high-resolution remote sensing image ground object segmentation,such as complex computation,low segmentation accuracy and large holes,this paper proposes DeepLabV3+network model with dual attention mechanism.This model uses lightweight MobilenetV2 as the backbone network,and fully considers three aspects of network parameter setting,edge extraction optimization and performance,so as to obtain dense context information.The model designed in this paper is applied to the high-resolution remote sensing image dataset of Xunwu county for verification.The results show that the DeepLabV3+network model with dual attention and resourceful achieves good segmentation accuracy for eight ground objects,especially for the segmentation of gardens,water bodies and roads,with the segmentation accuracy up to 92%,90%and 96%.This study provides a scientific reference for high-resolution image feature segmentation and how to make up for the defects of basic DeepLavV3+.
作者
龙北平
刘锟铭
占小芳
李恒凯
LONG Beiping;LIU Kunming;ZHAN Xiaofang;LI Hengkai(Geological Information Engineering Brigade,Jiangxi Geological Bureau,330001,Nanchang,PRC;Jiangxi University of Science and Technology,341000,Ganzhou,Jiangxi,PRC)
出处
《江西科学》
2023年第6期1093-1098,共6页
Jiangxi Science
基金
江西省地质局基金项目(360000228888030003254)。
