Unsupervised outdoor insulator defect detection based on multiscale feature reconstruction

Machine vision systems are widely used for outdoor insulator defect detection, crucial for preventing railway conveyor failures, yet challenging due to complex environments. Current methods depend on defect labels, while unsupervised models only achieve image-level classification and suffer from background interference. Our proposed multi-scale reconstruction model enables unsupervised defect detection by extracting fixed features through a feature extraction module, processing them via a multiscale approach to capture diverse defect sizes. An integrated attention module enhances feature capability, improving detection accuracy. Tests on a public outdoor insulator dataset show our algorithm surpasses mainstream unsupervised methods, achieving 94.8% pixel-level AUROC for precise defect localization.