摘要
深度学习是遥感影像建筑物提取的重要技术之一。针对目前卷积神经网络在提取建筑物时存在边缘模糊、不同尺寸建筑物提取结果差异大、模型参数量大等问题,提出一种基于注意力和多尺度特征融合的提取方法。利用高效通道注意力模块增强重要特征在网络训练的作用;嵌入多尺度特征深度融合模块提取并交互融合特征,同时借助卷积通道剪枝思想压缩模型。实验表明,该方法具有优异的建筑物提取能力,网络细节感知能力更强,提取边缘更加清晰,对复杂场景下不同尺寸和不规则建筑物的提取效果更好,并很好地平衡了模型提取精度和运行效率。
Deep learning is one of the important technologies for building extraction from remote sensing images.Aimed at the problems that the current convolutional neural network method has blurred edges,large differences in the extraction results of buildings of different sizes,and large amount of model parameters when extracting buildings,a new method based on attention and multi-scale feature fusion is proposed.The efficient channel attention module was used to enhance the role of important features in network training.The multi-scale feature deep fusion module was embedded to extract and interactively fuse features,and at the same time,the convolution channel pruning idea was used to compress the model.Experiments show that the method have excellent extraction ability,the detail-awareness of the network is stronger,the extracted edges are clearer,the extraction effect is better for buildings with different sizes and irregularities in complex scenes,and the model extraction accuracy and operation efficiency are well balanced.
作者
张琼
李百寿
夏金磊
张越
Zhang Qiong;Li Baishou;Xia Jinlei;Zhang Yue(College of Geomatics and Geoinformation,Guilin University of Technology,Guilin 541004,Guangxi,China;Guangxi Key Laboratory of Spatial Information and Geomatics,Guilin 541004,Guangxi,China;Guangxi Insititute of Serveying Mapping and Geoinformation,Nanning 530023,Guangxi,China)
出处
《计算机应用与软件》
北大核心
2025年第12期228-235,共8页
Computer Applications and Software
基金
国家自然科学基金项目(41161073)
桂林市科学技术与技术开发计划项目(20190601)。
关键词
遥感图像
语义分割
建筑物提取
通道注意力
多尺度特征
特征可视化
Remote sensing image
Semantic segmentation
Building extraction
Channel attention
Multi-scale features
Feature visualization
