Defect recognition of small samples in substations based on the domain adaptive YOLOv7
Jipan Li, Suliang Sun, Shuai Wang, Rui Guo, Yuheng Du, Yang Wang, Yuankai Han · 2023
The computer vision technology based on deep learning provides a new perspective on equipment inspection in power system. The intelligent technologies are now widely used in substations and definitely achieve positive effects, but they also face the problem of poor accuracy due to small samples. To improve the defect recognition in substations, this paper presents an improved YOLOv7 model that additionally includes the domain adaptation module, the SimAM attention module and the auxiliary detection head. The domain adaptation module strengthens the domain invariant features to improve the adaptability to unknown data. The SimAM attention module focuses on the key area and the objects. The auxiliary detection head can reduce the influence of large differences in object scale. The experiments on real-world data demonstrate that the improved model can better deal with the small samples and achieve better performance in the detection rate and the error rate.