Abstract: Metal corrosion will cause continuous damage to power equipment and form potential safety hazard. Detecting corrosion based on deep learning is presently becoming a feasible approach. The irregularity and detachability of metal corrosion makes annotators ambiguity and uncertainty in the labeling process. Accordingly, this paper proposes a novel hierarchical annotation method. Firstly, some large boxes are created, each of which labels a large area covering the range of corrosion. Secondly, in each labeling box created in the first step, regions with obvious corrosion features and relative independence are labeled to form the second layer of nested boxes. This method can easily produce unified annotation results without ambiguity. Finally, it highlights the corrosion features and increases the number of ground truth, realizing data augmentation and making it easier for detection models to learn the inherent features of corrosion during the training process. A series of experiments show that, through hierarchical annotation method, the recall of YOLOv5 increased from 50.79% to 59.41%, that of Faster R-CNN+VGG16 increased from 66.50% to 78.94%, and that of Faster R-CNN+Res101 increased from 78.32% to 84.61%.