Multiconstrained Heterogeneous Deep Network for Remote Sensing Rural Building Detection

Abstract

Remote sensing rural building detection holds substantial practical value for the scientific management and unified planning of rural land. However, most existing methods struggle to achieve desirable feature representations due to the similarities and imbalances between underconstruction buildings (UBs) and completed buildings (CBs), as well as interference from background noise, which results in high rates of false positives and false negatives. To address these issues, we propose multiconstrained heterogeneous deep network (MHDN) for remote sensing rural building detection. Specifically, we propose a grid-based CNN-GNN hybrid (GCGH) model that incorporates the sparse connectivity graph into the CNN backbone to model global feature correlations for more robust feature representations. Furthermore, a cross-image multiscale contrastive constraint (CMCC) branch is introduced to supervise network training alongside the detection loss, which facilitates detector learning in the presence of category imbalance. Experimental results on our proposed dataset demonstrate that our MHDN outperforms state-of-the-art object detection methods. The code and dataset are available at https://github.com/Dongxu-Wang/MHDN .