基于编解码网络UNet3+的遥感影像建筑变化检测  

Detection of Building Change in Remote Sensing Image Based on Encoder-Decoder Network UNet3+

作　　者：梁燕[1,2] 易春霞 王光宇 LIANG Yan;YI Chun-Xia;WANG Guang-Yu(School of Communication and In formatiom Engineering,Chongqing Universiry of Posts and Telecommunications,Chongqing 400065;Chongqing Key Laboratory of Signal and Information Processing,Chongqing 400065;Engineering Research Center of Mobile Communications of the Ministry of Educaion,Chongqing 400065)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]信号与信息处理重庆市重点实验室,重庆400065 [3]移动通信教育部工程研究中心,重庆400065

出　　处：《计算机学报》2023年第8期1720-1733,共14页Chinese Journal of Computers

基　　金：国家自然科学基金(61702066);重庆市教委科学技术重点研究项目基金(KJZD-M201900601)资助。

摘　　要：遥感影像建筑变化检测需解决两个重要问题:一是双时相影像本身存在的时间依赖性问题;其二是由于建筑物密集分布,阴影效应及各对象之间的相似性导致的特征鉴别问题.该文分析现有处理方案,提出了基于UNet3+网络的边缘引导变换检测网络(Edge-Guided Change Detection Base on UNet3+,EGCD-UNet3+).UNet3+利用全尺度的跳跃连接把来自不同尺度特征图中的深层语义与浅层语义直接结合,从多尺度聚合的特征图中学习层次表示,但是在特征提取时忽略了对象尺度规模,导致感受野与尺度不匹配.因此EGCD-UNet3+首先设计了一种具有自适应感受野的选择性核Block(Selective Kernel Block,SKB)代替UNet3+原始的Block,使影像对在提取深、浅层特征时具有自适应感受野属性.EGCD-UNet3+由编码与解码两部分构成,在编码端利用长短期记忆网络(Long Short-Term Memory,LSTM)捕捉长期依赖关系,建模像素之间的关系上下文,设计差分增强模块(Difference Enhance Module,DEM),分析影像对之间的时间相关性,解决双时相本身存在的时间依赖性问题.在解码端,EGCD-UNet3+提出边缘引导上下文模块(Edge-Guided Context Module,EGCM)进一步改善建筑检测边界的性能,在更细粒度水平上有效提取多尺度空间边缘信息.最后,EGCD-UNet3+利用同时具备像素分割误差和边缘分割误差的复合损失函数,使网络能够充分学习有效的特征进行准确的标签预测.所提模型在LEVIR-CD与WHU-CD数据集上验证,精准率(P)分别达到90.75%、91.75%,召回率(R)可分别增长到96.68%、92.42%,F1-score(F1)分别增加到93.15%、92.08%,总体分割精确度(OA)分别达到99.12%、98.96%,且交并比(IoU)分别增加到83.96%、74.91%.There are two important problems in building change detection of remote sensing image:one is the time dependence problem of the bi-temporal image itself,the other is the object identification problem due to the dense distribution of buildings,shadow effect and the similarity between the features.This paper proposes an Edge-Guided Change Detection Base on UNet3+(EGCD-UNet3+)network after analyzing the existing schemes.The UNet3+uses the full-scale skip connections to directly combine the deep and shallow semantics from different scale feature maps,and learns the hierarchical representation from the multi-scale aggregated feature maps.However,it ignores the object scale in feature extraction,leading to the mismatch between the receptive field and the scale.Therefore,the EGCD-UNet3+first designs a Selective Kernel Block(SKB)with adaptive receptive field to replace the original UNet3+block,so that the image of deep and shallow features is extracted with adaptive receptive field properties.The EGCD-UNet3+consists of the encoding part and the decoding part.In the encoding part,the Long Short Term Memory(LSTM)is used to design a Difference Enhancement Module(DEM)for analyzing the temporal correlation between image pairs,and solving the time dependency problem of the bi-temporal itself,because the LSTM's abilities of capturing long-term dependency,modeling the context relationship.In the decoding part,the EGCD-UNet3+proposes an Edge Guided Context Module(EGCM)to further improve the performance of building detection boundary,and effectively extract multi-scale spatial edge information at a finer-grained level.Finally,the EGCD-UNet3+defines the composite loss function with pixel segmentation error and edge segmentation error,so that the network can fully learn effective features for accurate label prediction.The proposed model has been experimentally verified on the LEVIR-CD and WHU-CD dataset.The results show that the precision(P)reached 90.75%and 91.75%respec-tively,the recall(R)increased to 96.68%and 92.42%respectively,

关 键 词：变化检测 差分增强 长短期记忆 选择性核 边缘引导上下文 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]