An Active Learning Model Production Method for Electric Power Inspection Image Multi-target Detection
Jiangqi Chen, Bo Wang, Xi Zhang, Kunlun Gao, Peng Wu · 2023
With the development of intelligent electric power inspection technology, electric power companies use UAVs, cameras, robots and other equipment for power transmission line inspections, and take a large number of inspection images. In recent years, the image multi-target detection model has been maturely applied in the intelligent inspection of electric power. The current image intelligent detection model mainly relies on a large number of image sample labels. Power companies have a small number of labeled image samples, and a large number of unlabeled. But manual data annotation is very expensive, and cost lots of time and money. Moreover, there will be certain errors in manual annotation, so it is necessary to repeat the annotation and verification to reduce the error. But this will also increase the cost. This article proposes an active learning model production method for electric power inspection image multi-target detection. We select samples for annotation according to some strategies, and select labeled samples for verification. This method can reduce the cost of sample annotation, and improve the model. productivity. Compared with the traditional random labeling strategy, our method can achieve the same accuracy with only 20% of the annotation cost. With 50% of the annotation cost, the mAP index will increase by 12%.