MBCNet: Multi-Branch Collaborative Change-Detection Network Based on Siamese Structure

The change-detection task is essentially a binary semantic segmentation of changing and invariant regions. However, this much more difficult than simple tasks, as the areas typically include multiple terrains such factories, farmland, roads, buildings, mining areas. This requires ability network to extract features. To end, we propose multi-branch collaborative based on Siamese structure (MHCNet). In model, three branches, difference branch, global similar are constructed refine information from remote-sensing images. Four modules, cross-scale feature-attention module (CSAM), filtering (GSFM), double-branch information-fusion (DBIFM), similarity-enhancement (SEM), proposed assist branches better. CSFM used related change in image GSFM filter rich image, DBIFM fuse extracted branch branch. Finally, SEM uses with correct details feature map feature-recovery stage.