The Low-Illumination Catenary Component Detection Model Based on Semi-Supervised Learning and Adversarial Domain Adaptation

Catenary component status detection is a core task of high-speed railway infrastructure operation and maintenance. Since catenary detection is mainly carried out during the nighttime railway skylight period, the illumination of some collected catenary images is lower, which is often ignored when labeling data, which reduces the detection accuracy and robustness of component damage identification methods. To this end, this article proposes a low-illumination catenary component detection model based on semi-supervision and adversarial domain adaptation. First, a neighbor mask consistency semi-supervised (NMCS) method is proposed. According to the sight occlusion changes of the catenary structure, a neighbor mask block coverage strategy is designed and integrated into the semi-supervised teacher-student model to effectively improve the component recognition accuracy of the proposed model. Second, a dynamic weight feature domain adaptive alignment (DWFDA) method is proposed. This method introduces adversarial learning and uses the A-distance to construct a dynamic weight multiscale global image feature and local instance feature adaptive alignment model to achieve adaptive alignment of low-illumination image features to normal illumination image features, further improving the detection accuracy of the proposed model. The results show that the proposed model can achieve accurate component detection of images with different illumination levels without increasing the parameters of the original detection model, making the detection accuracy of low-illumination unlabeled images 88.71% and reaching a nearly consistent level with the detection accuracy of normal illumination labeled images 89.23%, improving the efficiency and robustness of the detection model.