General Visual Prior Integrated YOLOv8 for Accurate Insulator Detection of Power Distribution Networks in Complex Environments

Detecting distribution network insulators with deep learning algorithms is challenging due to the complex background and the small size of the targets. Since Visual Foundation Models(VFM) have shown impressive performance in general domains, this paper attempts to utilize the features extracted from VFM to assist YOLOv8 in achieving more accurate detection. InternImage is selected to provide the general visual priors, and its features are combined with the feature maps of YOLOv8 after passing through a feature fusion block. The proposed model is evaluated through extensive experiments by training on the distribution network insulator dataset and demonstrated to be superior to the original YOLOv8 in this task. Further, the superiority of the proposed model in detecting small-size insulators is illustrated by visualization. Additionally, the effectiveness of the proposed solution in identifying small-size insulators is demonstrated through visualization.