Enhanced YOLOv9s‐Based Surface Defect Detection in Filter Images via Generative Data Augmentation and Model Optimisation

ABSTRACT Surface defect detection in oil filters is crucial for maintaining engine performance and longevity. However, the scarcity of defect datasets and the challenges in detecting minor surface defects have hindered the effectiveness of existing deep learning models. To address these issues, we propose a dataset generation method based on StyleGAN3 and a filter surface defect detection algorithm, SCP‐YOLO, based on an improved YOLOv9s. By generating filter defect images using StyleGAN3 and combining them with filter images, we create a large‐scale dataset. The CloBlock module is incorporated into YOLOv9s to enhance the model's focus on small target defects, while the Spatial and Channel Reconstruction Convolution module is adopted to create a more efficient feature extraction backbone, balancing the computational cost introduced by the attention mechanism. Additionally, the Powerful Intersection over Union loss function is used to improve convergence speed and model accuracy. Experimental results show that SCP‐YOLO achieves an mAP@0.5 of 97.34% and an mAP@0.5:0.95 of 77.42%, with a frame rate of 173 f/s, demonstrating its effectiveness in real‐time filter surface defect detection.