实验技术与管理2025，Vol.42Issue(4)：48-58,11.DOI:10.16791/j.cnki.sjg.2025.04.007
用于遥感图像变化检测的交互式多编码器和多解码器网络设计
IMEMD-Net:An interactive multiencoder and multidecoder network for enhanced change detection in remote sensing images
摘要
Abstract
[Objective]The multiencoder and single decoder(MESD)architecture has demonstrated success in change detection tasks.This method employs Siamese encoders to independently extract features from bitemporal images,effectively preserving the target features of each temporal image and enhancing the ability to capture changes in input data.Typically,MESD architectures use convolutional neural networks(CNNs)or transformers as encoders.CNNs are highly effective in capturing complex and contextually relevant features,and they can extract local features effectively.However,CNNs struggle to capture global dependencies.Conversely,transformers,with their self-attention mechanism,excel at handling complex spatial transformations and capturing long-range feature dependencies,thereby providing robust global representation.However,transformers may occasionally overlook local feature details,which reduces the clarity of foreground and background elements.This limitation can propagate through the fusion process,resulting in blurred object boundaries and false changes in the final change map.Using CNNs or transformers separately as encoders presents challenges in terms of balancing global and local feature extraction,potentially hampering feature representation.In addition,most MESD methods employ a single decoder that can fully leverage the diverse and rich information provided by multiple encoders,thereby influencing the accuracy and precision of the change detection results.To address these issues,this study proposes an interactive multiencoder and multidecoder network(IMEMD-Net).[Methods]The proposed IMEMD-Net employs Siamese interactive encoders to extract local and global features from bitemporal images.The interactive encoder comprises a parallel structure that combines CNNs and transformers to maximize the retention of local features and global representations.The local features extracted by CNNs and the global features extracted by transformers are enhanced using a feature communication module(FCM),which continuously resolves semantic discrepancies between the two.The fusion process significantly improves the global perception of local features and integrates local details into global representations.Next,global and local difference decoders process the multiscale local and global features extracted from the bitemporal images,with each decoder focusing on specific types of difference features.This targeted processing reduces the computational burden and improves overall performance.Finally,a spatial channel difference fusion(SCDF)module adaptively fuses the difference features obtained from both decoders across spatial and channel dimensions,effectively enhancing relevant changes while suppressing false positives.[Results]The algorithm's performance is evaluated using metrics such as precision,recall,F1-score,intersection over union(IoU),and overall accuracy.Compared to existing state-of-the-art MESD-based methods,such as ChangeForm,the proposed algorithm achieves F1-score improvements of 1.14%,5.68%,and 6.5%,along with IoU score increases of 1.92%,9.66%,and 10.73%on the LEVIR-CD,WHU-CD,and DESIF-CD datasets,respectively—without requiring pretrained backbone feature extractors.[Conclusions]This approach effectively integrates CNNs and transformers through the FCM,enhancing interaction to ensure the extraction of global and local features from bitemporal images.The use of dual decoders alleviates the limitations of single-decoder architectures by distributing the task of processing local and global difference features.In addition,the SCDF module improves the method's ability to extract more discriminative change features,which significantly enhances the efficiency of change detection in remote sensing images.关键词
遥感图像/变化检测/Transformer/CNNKey words
remote sensing/change detection/Transformer/convolutional neural networks分类
信息技术与安全科学引用本文复制引用
王雷全,童寿梁,耿辰东..用于遥感图像变化检测的交互式多编码器和多解码器网络设计[J].实验技术与管理,2025,42(4):48-58,11.基金项目
中国石油大学(华东)研究生教育教学改革项目(YJG20230044) （华东）
中国高等教育学会"校企合作双百计划"典型案例(GJXH-SBJHDX-2022157) （GJXH-SBJHDX-2022157）
国家自然科学基金项目(62071491) （62071491）
山东省本科教学改革研究重点项目(Z2023006) （Z2023006）
山东省研究生精品和优质课程(SDJKC2003008) （SDJKC2003008）
中国石油大学(华东)研究生精品示范课程建设项目(UPCYJP-2023-07) （华东）
