MFIHNet: Multi-Scale Feature Interaction Hybrid Network for Change Detection of Remote Sensing Images
Lin Cao,Qi Liu,Shu Tian,Lihong Kang,Jing Tian,Xiangwei Xing,Kangning Du,Huanyu Bian,Peiran Song,Yanan Guo,Chunzhuo Fan,Chong Fu,Ye Zhang
DOI: https://doi.org/10.1109/jstars.2024.3450287
2024-01-01
Abstract:Remote sensing image change detection (RSCD) based on deep learning technology has made remarkable achievements. Meanwhile, the enhancement of network architectures and advancements in optimization algorithms have pushed RSCD performance to a higher stage. However, the existing RSCD methods mainly focus on extracting differential information between pixel pairs from bitemporal images, while neglecting the importance of using complementary multiscale features to uncover hierarchical semantic change information. To address these issues, we propose a multiscale feature interaction hybrid network (MFIHNet), which aims to enhance feature discriminability by multiscale features interaction and fusion to enhance model RSCD performance. Specifically, we first design a cascaded convolutional neural networks (CNN)-Transformer feature extraction network to capture hierarchical features at different scales. This strategy enables the network to preserve detailed information in shallow layers while grasping more contextual information in high layers. Subsequently, based on the differences between hierarchical features, we design a novel edge enhancement module (EM) to adaptively focus on key areas under the guidance of edge information to make the changed information clearer. Furthermore, to ensure complementary advantages among different feature layers, we devise a novel cross-scale feature interaction module, which introduces a region-specific atrous convolution into the multiscale attention mechanism for improving feature coassistance capacity. In this way, the MFIHNet not only effectively obtains different types of fine-grained information but also reduces the loss incurred during the feature fusion process, thereby improving the performance of remote sensing RSCD tasks. Extensive experimental results on the challenging CDD and GZ-CD datasets, with mean F1 scores reaching 98.1% and 87.4%, respectively, demonstrate that the proposed method achieves competitive performance.