Small Data-Driven Electrical Insulator Defect Detection

The inspection and maintenance of insulator equipment has always adopted the traditional manual detection. It is very significant to study the automatic Insulator defect detection by drone inspection. However, in practical industrial applications, the number of available defect insulator samples is limited. It is difficult to construct a sufficient and high-quality dataset to support the training of the object detection model. In this paper, we propose a detection framework which combines the super-resolution reconstruction and the object detection model. In our model, we use the super-resolution reconstruction and traditional data augmentation to amplify the amount of data and avoid the overfitting caused by the small sample data. The model has excellent performance on the training set which only contains 80 images, and achieves 61% mAP. We also show that the super-resolution reconstruction can rich image texture features and is more effective than some traditional data augmentation methods.