A cascading YOLOx model of fitting detection for high-resolution aerial image

Abstract

The fittings detection of transmission lines plays a vital role in ensuring the safe and stable operation of transmission lines. The fitting detection method based on deep learning only scales the original image to a smaller size. However, it ignores the high-definition resolution of the aerial image in the transmission line, resulting in the loss of rich features in the high-resolution aerial image. In order to solve this problem, we observe that aerial images of fittings are concentrated in a particular area of the aerial image. Therefore, we propose a cascading YOLOx model, including Dense Target Regions YOLOx (DTR-YOLOx), which can detect dense target areas, and Multi-Fitting YOLOx (MF-YOLOx), which can detect multiple categories of fittings. In addition, an algorithm based on the connected regions is proposed to automatically generate the dense target region for DTR-YOLOx training, reducing manual labeling costs. Furthermore, the EIOU loss function is introduced to improve the precision of the model's coordinate regression. Experiments show that the AP 0.50:0.95 value of our proposed model is 18.4% higher than that of the YOLOx model.