Overhead Power Line Detection Based on Curriculum Learning Strategy

Drone inspection of overhead power lines is a crucial aspect of ensuring electrical safety. Due to the elongated shape of overhead power lines and the complex inspection backgrounds, commonly used detection networks struggle to learn the unique features of these targets. To achieve high-precision detection of overhead power lines, a detection method based on a curriculum learning strategy is proposed. First, multiple target bounding boxes are annotated for individual overhead power lines in aerial images. Next, simulating the human learning process of starting with easier tasks and progressing to more difficult ones, the wavelet-entropy algorithm is used to divide the dataset into simple and complex subsets, allowing the YOLOv5 object detection network to first learn from the simple data and then from the complex data. Finally, a false detection box removal algorithm based on directional gradient statistics is introduced to process the overhead power line images that still contain incorrect identifications after network detection, eliminating the erroneous detection boxes. Experimental results indicate that, under the same network scale, high-precision detection of overhead power lines has been achieved, with an accuracy improvement of 10.8% compared to YOLOv5 based on traditional training strategies, and mAP_0.5 reaching 78.6%. This method effectively detects slender targets like overhead power lines while minimizing background interference, advancing the development of drone-based electrical inspection technology.