PowerDet: Efficient and Lightweight Object Detection for Electric Power Open Scenes

Abstract

In recent years, with the expansion of the electric network, maintenance cost of electric power transmission and transformer substation equipment has become increasingly greater. Moreover, it requires a lot of manpower as well. At present, there are certain schemes which leverage artificial intelligence techniques to detect equipment flaws or errors automatically. Nevertheless, there are problems with the schemes in open electric power scenes, e.g., only able to detect single category, low detection accuracy of multi-scale objects, and difficulty in deploying models on mobile devices. To address the challenges mentioned above, we propose an object detection model, named PowerDet. It is able to detect 9 different types of power facilities efficiently with low cost. To verify the effectiveness of PowerDet, we collect an open scene facility entity dataset and conduct a series of experiments. Experimental results demonstrate that PowerDet achieves 86.8% AP50 on the dataset, which outperforms the state-of-the-art. The lightweight version of PowerDet, i.e., PowerDet-Lite, can achieve real-time inference on mainstream mobile devices.