摘要
针对在不同地物背景下遥感影像道路提取中仍存在大量漏提、误提、提取精度不够高等问题,文章基于U-Net模型构建了Res50CBAM-Net模型。首先,该模型将原始U-Net模型的特征提取网络替换为ResNet50,加深了特征提取网络深度,提高了网络特征提取能力;其次,在U-Net跳跃连接层加入卷积块注意力机制,增强了模型对道路目标的识别能力。结果表明,文章构建的模型在不同场景下提取效果更好,交并比、F1-score较原始模型分别提升了2.72%、2.26%。
In response to the problems of significant omissions,errors,and low extraction accuracy in road extraction from rcmote sensing images under different land cover backgrounds,this paper constructs a Res5OCBAM Net model based on the U-Net model.Firstly,the model rcplaces the feature extraction network of the original U-Net model with ResNet50,deepening the depth of the feature extraction network and improving its feature extraction capability;Secondly,the addition of convolutional block attention mechanism in the U-Nct skip connection layer enhances the model's ability to recognize road targets.The results show that the model constructed in this article has better extraction performance in different scenarios,with an interscction to union ratio and F1 score improvemcnt of 2.72%and 2.26%,respectively,compared to the original model.
作者
苏雷
屈炳剑
SU Lei;QU Bingjian(Meitan County Land space Planning Ecological Restoration Engineering Technology Center,Zunyi,Guizhou 563000;Meitan County Natural Resources Survey and Land spatial Planning Center,Zunyi,Guizhou 563000)
出处
《长江信息通信》
2024年第3期83-85,共3页
Changjiang Information & Communications