Semisupervised Cross Domain Teacher-Student Mutual Training for Damaged Building Detection

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