摘要
卷积神经网络(Convolutional Neural Network,CNN)和Transformer的混合架构能够有效建模图像的局部与全局特征,已成为遥感图像变化检测任务的主流网络.然而这类网络仍面临着一些挑战. CNN分支中的卷积和池化运算通常会抑制遥感图像中的高频信息,降低目标边界的精度;此外,Transformer分支对图像像素进行等同长程依赖关系建模,忽略了变化目标的形状及语义关联信息,导致网络对变化目标特征的表达不足.为解决上述问题,提出了基于边缘引导和动态可变形Transformer的遥感图像变化检测网络.在CNN分支中设计了边缘信息引导模块,利用高频信息增强目标区域的边缘信息,从而改善变化目标的轮廓精度.同时设计了一种新颖的动态可变形Transformer,能够自适应地匹配形状不同的变化目标,选择与变化相关的特征建模长程依赖关系,以提高网络的特征表达能力.实验结果表明,提出的方法在三个公开数据集LEVIR-CD、CDD和DSIFN-CD上显著提高了检测精度,在变化目标的边界精度和内部完整性方面都明显优于当前的主流网络.
The hybrid architecture of convolutional neural network(CNN)and Transformer can effectively model lo-cal and global features of images,and has emerged as the predominant choice for remote sensing images change detection tasks.Nevertheless,these networks still confront challenges.The convolution and pooling operations employed by the CNN branch typically suppress the high-frequency information of remote sensing images,resulting in decreased precision of object boundaries in change detection results.Additionally,the Transformer branch equivalently models long-range de-pendencies for all pixels in remote sensing images,thereby disregarding shape information and semantic associations of ob-jects,which limits the network’s feature representation ability on changed objects in remote sensing images.To address these challenges,a remote sensing images change detection network is proposed based on edge guidance and dynamic de-formable Transformer.In the CNN branch,an edge information guidance module(EIG)is designed to enhance the edge in-formation of changed objects by leveraging the high-frequency details of images.This enhancement improves the edge ac-curacy of the changed objects.Simultaneously,an innovative dynamically deformable Transformer(DDaT)is designed to adaptively match changed objects with different shapes,selecting features relevant to changes to model long-range depen-dency relationships and enhance the network’s feature expression capability.Experimental results show that the pro-posed method significantly improves the detection accuracy on three public datasets:LEVIR-CD,CDD and DSIFN-CD,and is significantly better than the current mainstream networks in terms of edge accuracy and internal integrity of changed objects.
作者
雷涛
翟钰杰
许叶彤
王营博
公茂果
LEI Tao;ZHAI Yu-jie;XU Ye-tong;WANG Ying-bo;GONG Mao-guo(Shaanxi Joint Laboratory of Artificial Intelligence,Shaanxi University of Science and Technology,Xi’an,Shaanxi 710021,China;School of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi’an,Shaanxi 710021,China;Key Laboratory of Collaborative intelligent Systems,Ministry of Education,Xidian University,Xi’an,Shaanxi 710071,China)
出处
《电子学报》
EI
CAS
CSCD
北大核心
2024年第1期107-117,共11页
Acta Electronica Sinica
基金
国家自然科学基金(No.62271296,No.62201334)
陕西省杰出青年科学基金(No.2021JC-47)
陕西省重点研发计划(No.2022GY-436,No.2021ZDLGY08-07)
陕西省自然科学基础研究计划(No.2022JQ-634,No.2022JQ-018)
陕西省创新能力支撑计划(No.2020SS-03)。
