Feature enhancement YOLO for high-precision infrared thermography in non-destructive testing

Infrared thermography in non-destructive testing has been widely applied in sectors such as power equipment, aerospace, and materials science due to its advantages of non-contact operation, real-time performance, and high efficiency. However, infrared images suffer from issues like low resolution, noise interference, and blurred target edges, particularly when detecting small defects. Traditional detection methods often struggle to achieve accurate identification under these conditions. To address this, this paper proposes an improved YOLO algorithm with Semantic-Optimized Feature Enhancement Module (SOFEM-YOLO), which integrates a small object tracking mechanism into the original YOLO framework. By incorporating a feature enhancement module and tracking mechanism for small targets, the proposed algorithm significantly improves detection accuracy and stability for small infrared defects. Experimental results demonstrate that SOFEM-YOLO outperforms traditional YOLO backbone model in detecting power equipment flashovers, bubbles, impurities, and other defects.