摘要
针对现有变化检测方法中单分支网络将原始影像直接求和造成的差异信息损失以及中小目标误检等问题,提出一种基于U-Net的结合差异增强的变化检测网络。将差分图像经过通道注意力机制学习每个通道在特征表示中的差异性,生成与通道相关的权重;利用这些权重与原始图像加权求和,然后将特征增强后的两期图像融合后输入网络;接着通过密集残差块加强编码器信息传递和复用;利用不同形式和尺度的卷积进一步细化检测结果。所提方法在Sardinia和Shuguang数据集中与主流方法进行对比,相较于MUNet,OA分别提升了1.27%和0.74%;F1分别提升了5.32%和1.75%。结果表明,所提方法能够充分利用差异信息,对中小目标以及边缘细节有较高的分割能力。
The purpose of this study is to address the issues of loss of differential information and false detections of small and medium-sized targets caused by the direct summation of the original images in the existing change detection methods using single-branch networks.A change detection network based on U-Net with binding difference enhancement was proposed.First,the differential image is passed through a channel attention mechanism to learn the distinctiveness of each channel in feature representation,generating weights that are relevant to each channel.Furthermore,the weights obtained are used to perform a weighted sum with the original images,and then the two enhanced images are merged to be fed into the network.Then,the dense residual block is used to strengthen the encoder information transmission and reuse.Finally,the detection results are further refined using convolutions of different forms and scales.Compared with the mainstream methods in the Sardinia and Shuguang datasets,the OA of the proposed method was increased by 1.27 and 0.74 percentage points respectively,compared with the MUNet,F1 was improved by 5.32 and 1.75 percentage points respectively.The conclusion is that the proposed method can fully utilize the differential information and has high segmentation ability for small and medium targets as well as edge details.
作者
邱云飞
屈照阳
方立
QIU Yunfei;QU Zhaoyang;FANG Li(College of Software,Liaoning Technical University,Huludao,Liaoning 125105,China;Laboratory of Remote Sensing and Information Engineering,Quanzhou Institute of Equipment Manufacturing,Haixi Research Institute,Chinese Academy of Sciences,Quanzhou,Fujian 362216,China)
出处
《测绘科学》
CSCD
北大核心
2024年第1期153-162,共10页
Science of Surveying and Mapping
基金
国家自然科学基金项目(62173171)
国家自然科学基金-青年基金项目(42101359)
辽宁省自然科学基金资助项目(2015020095)
辽宁省教育厅科学技术研究资助项目(LJYL051)
关键词
异源遥感影像
变化检测
密集残差
差异增强
空洞多尺度卷积
heterogeneous remote sensing images
change detection
dense residuals
,differential augmentation
atrous multi-scale convolution
作者简介
邱云飞(1976—),男,辽宁阜新人,教授,主要研究方向为机器视觉、图像处理等。E-mail:7415575@qq.com;通信作者:屈照阳,硕士研究生,E-mail:674480687@qq.com
