Surface defect detection method of aluminium profile based on improved YOLOv8

In aluminum profile production, surface defects like scratches and bottom exposure reduce product quality and lifespan. Traditional manual and single-signal detection methods are limited. This paper proposes a defect detection method for aluminum profile based on improved YOLOv8. The SimAM module is integrated into the YOLOv8s backbone, and the CA attention mechanism is added to the neck structure, enhancing feature information flow and narrow feature map representation. Ghost convolution replaces standard convolution, reducing model parameters and computational complexity. To address the imbalance in defect aspect ratios, a width height balance factor is introduced into the EIOU loss function, improving bounding box regression accuracy. Dynamic adjustment of mosaic parameters further optimizes training, enhancing model generalization and detection performance. Experimental results show a 4.84% increase in F1 score and a 14.88% jump in mAP@50:95, demonstrating the model’s superiority over others.