An efficient image enhancement method for transformer internal defect recognition via autonomous robotic fish

Currently, detecting the internal status of large transformers often involves labour-intensive methods like manual core drilling or hanging cover inspection. To promote the use of robot fish to replace manual inspection of the operating conditions inside the transformer, this paper designs a transformer internal image enhancement, defect detection and segmentation system, and explores the feasibility of robot fish replacing manual inspection. Initially, a self-calibrated illumination network is combined with histogram equalisation to improve the internal image quality of the transformer. Then, an image super-resolution network is introduced to restore the lost details in low-resolution transformer interior images, enabling technicians to better judge the operating status of the transformer through the images taken by the robot fish. Furthermore, we study the possible defect types inside the transformer, then build and augment the existing defect database. Subsequently, transformer internal defects are detected through the state-of-the-art YOLOv10 model, and SAM is introduced to perform instance-level segmentation on images within the defect object box. Experimental results demonstrate that our method significantly enhances the visual quality and resolution of transformer internal images and excels in detecting defects, achieving a mAP 50 of 98% on the augmented defect dataset.