Pre-Trained Large Model Based Automatic Generation of Defect Insulator Images for Power Distribution Systems

The imbalance between normal and defect insulator samples brings biases to the neural network, misleading to overfitting and misclassification. To tackle this problem, we propose an image generation method leveraging pre-trained Large Models (LMs) to augment defect samples, especially cap-exploded insulators. Specifically, we devise the cap segment module to generate semantic mask covering insulator caps. Subsequently, the inpainting module repairs the blocked regions with the background, creating cap-explosion-like defect samples. We investigate the optimal injection rate of pseudo images and assess the generation quality. Learned Perceptual Image Patch Similarity (LPIPS) is utilized to select useful generated samples to retrain and enhance performance. Comparative experiments are conducted on a dataset comprising 1203 normal insulators and 383 defect ones. Our method improves detection F1 scores up to 7% and 21 % on ResNet18 and ViT-B/16, respectively.