Semi-Supervised Cross Domain Teacher-Student Mutual Training for Damaged Building Detection

Detection of damaged buildings is a form object detection and essential for disaster emergency response efforts. In recent years, deep learning has been widely used in detection, with successful target models such as Faster-Rcnn YOLO being proposed. However, training usually requires large amount labeled data. Due to the high threshold aerial remote sensing data collection, collapsed very sparse. addition, limited area damage single scene leads insufficient feature diversity, which can easily lead model overfitting. These issues restrict development applications. To solve these problems, we propose paradigm named cross-domain teacher-student mutual training. By using Cycle-GAN generated style transfer through teacher network, pseudo-labels are train student network. Then, network slowly updates parameters indirectly learn generalization information satellite domain. Networks trained this way achieve good results detecting houses aviation We tested on our self-built dataset, DB-ARSD, includes bounding box labeling buildings, found that method outperforms other methods both house prediction accuracy domain performance.