Abstract: The safety of high-voltage power line is usually threatened seriously by the foreign matters such as bird’s nests or kites and so on. The transmission tower is an important part of the system of high-voltage power line. So the UAV (Unmanned Aerial Vehicle) often takes photos of tower especially when inspecting the high-voltage power line and these photos usually have to be analyzed by the classical algorithm of YOLOv3 (You Only Look Once Version 3) whether they contain the foreign matters or not. This research is conducted to improve the classical algorithm of YOLOv3 in order to improve its precision and deficiency and unknown scale of weight parameters. Initially, the structure of improvement is designed and image data set is constructed. Second, the classical algorithm of YOLOv3 is improved from three ways: the width and height of loss functions of prediction box, the unbalanced loss function of prediction type and the network structure of classical algorithm are improved respectively. The experiments show that: the improvements proposed are effective, which improves the average recognition precision and reduces the scale of weight parameters greatly and maintains the efficiency. The result of improving the classical algorithm’s neural network structure is obviously better than other ways, which is probably the main direction to improving the algorithm. The investigation provides important basics to detect the objects on real time of UAV.

Key words: high-voltage power transmission line inspection, image detection, bird's nest recognition, YOLOv3 algorithm, neural network