Insulators Classification with Optimized DETR Models and Spatial-Adaptive Networks

To ensure the reliability of electrical power supply, the electric utility conducts thorough inspections of the power grid. In modern times, unmanned aerial vehicles are utilized for their adeptness in capturing images in challenging or inaccessible locations. However, accurately classifying insulator conditions during these inspections using computer vision presents a significant challenge, primarily due to the potential performance limitations of standard object identification methods. Typically, these methods are pre-trained for generalized tasks and may not be specialized for insulators. This paper introduces an innovative framework for insulator classification, which consists of an optimized DETR-like detection model and a spatial-adaptive decision network. Initially, we enhance the general pre-trained DETR model, named DINO, by incorporating domain knowledge obtained from the target datasets. We employ a dynamic re-sampling strategy for the training data to ensure effective detection performance specifically tailored to the target category, while also distilling detection capabilities from other open-world categories. Acknowledging the necessity for varied receptive fields due to the diverse shapes of insulators, our decision network integrates adaptive Convolutional Neural Network (CNN) and Transformer fusion structures. Extensive experiments conducted on representative benchmarks validate the effectiveness of our proposed framework.