A Visual Foundation Model of Image Segmentation for Power Systems

The automation of substation equipment inspection is a pivotal development area within the power industry. Traditional substation equipment inspection methods utilizing instance segmentation models trained on specific dataset have shown broad application, however, their generalization performance is limited to specific scenes. To enhance the robustness against intricate environments, we propose a two-stage instance segmentation method based on visual foundation models. In our work, the state-of-the-art object detector YOLOX and the visual foundation model SAM are employed to integrate the high-efficiency 2D detector with the general visual knowledge powered by foundation models trained on large-scale datasets. We utilize YOLOX to generate bounding box prompts which are processed by the pruned and aligned visual foundation model SlimSAM to perform instance segmentation with 68.6% mAP on our validation dataset. The method's effectiveness is validated through extensive comparisons with different model configurations and segmentation prompts, highlighting its robustness and potential for practical application in the domain of substation equipment maintenance and inspection.