YOLO-Based Component Detection System for Electrical Diagrams

In power grid engineering, improving work efficiency is crucial, and accurate recognition of power grid diagrams is particularly important. To address issues such as the insufficient availability of power grid diagram datasets and the limited accuracy in detecting electrical symbols, we proposed a method involving the creation of a custom dataset for comparative experiments and the addition of post-processing operations. Geometric transformations were used for data augmentation to ensure an adequate number of training samples. The custom dataset was used to train the YOLOv5s, YOLOv10s, and FCOS models. Comparative experimental results showed that YOLOv10s achieved higher precision and mAP compared to YOLOv5s and FCOS, with increases of 3.9 and 2.5 percentage points, respectively, and an increase of 8.4 units in FLOPs, indicating superior performance of YOLOv10s. Additionally, the incorporation of HOG feature-based template matching in the post-processing stage further reduced the false positive rate, significantly enhancing detection accuracy.