MBNet: Multi-Branch Network for Extraction of Rural Homesteads Based on Aerial Images

Deep convolution neural network (DCNN) technology has achieved great success in extracting buildings from aerial images. However, the current mainstream algorithms are not satisfactory feature extraction and classification of homesteads, especially complex rural scenarios. This study proposes a deep convolutional for homestead consisting detail branch, semantic boundary namely Multi-Branch Network (MBNet). Meanwhile, multi-task joint loss function is designed to constrain consistency bounds masks with their respective labels. Specifically, MBNet guarantees details prediction through serial 4× down-sampled high-resolution maps adds mixed-scale spatial attention module at tail branch obtain multi-scale affinity features. At same time, low-resolution interaction between maintained. Finally, result segmentation refined by point-to-point (PTPM) generated boundary. Experiments on UAV imagery areas show that our method achieves better performance than other state-of-the-art models, which helps refine homesteads. demonstrates potential candidate building an automatic management system.