Application of a Small-scale Object Detection Method Based on Model Cascade in Inspection of Transmission Lines by Unmanned Aerial Vehicles

The bolt pin serves as an essential part of preventing the nut from slipping in overhead transmission lines, with its loss potentially leading to serious transmission safety accidents. In this regard, unmanned aerial vehicle (UAV) inspection can be employed to effectively determine such defects. The manual inspection of aerial images, however, inevitably exposes some defects, such as low efficiency, huge workload, and high repetition. While the traditional object detection method can be utilized to detect apparent defects, the detection effect is not ideal due to the excessively small pin scale and interference from complex backgrounds and similar parts. To this end, this paper proposes a small-scale object detection method based on model cascade, which automatically detects the loss of bolt pins in transmission lines by training two detectors. Specifically, the primary detector is leveraged to detect the object area and eliminate the interference of similar objects. In contrast, the secondary detector performs the detection of small-scale objects based on the primary detection results. Experimental results indicate that the AP50 and mAP50 of the proposed method have seen significant improvement on detection of normal and abnormal objects compared with those of single-stage detector YOLOv5, multi-stage detectors Faster RCNN and Cascade R-CNN. Simply put, the method proposed in this paper can effectively improve the detection accuracy of small-scale objects by leveraging domain knowledge to design the detection process" and enhance the efficiency and safety of inspections, thereby providing valuable inspiration for its practical application in related fields.