Super-resolution-based data augmentation algorithm for few-shot bearing defect detection

Accurate detection of bearing surface defects is essential for ensuring operational efficiency and extending the service life of mechanical equipment in engineering and industrial manufacturing. However, defect detection performance is often limited by few data in high-dimensional spaces. Traditional VAE architectures, while capable of data augmentation, frequently produce images with insufficient detail and clarity. This paper proposes a super-resolution-based data augmentation method that utilizes Swin Transformer modules for feature extraction and processing, combined with CARAFE and Coordattention mechanisms for upsampling, to enhance image quality and defect detail in few-shot bearing data. Additionally, perceptual and texture losses are introduced to further optimize visual quality. Experimental results show that the proposed method significantly improves image clarity and detail across multiple datasets, demonstrating strong robustness in few-shot scenarios and offering an effective solution for mechanical defect detection and high-dimensional few-shot learning.