Abstract Additive manufacturing (AM) is transforming industrial production; however, inevitable defects—such as spaghetti-like collapses, surface blemishes (“zits”), and stringing—substantially degrade product quality and mechanical performance. To overcome limitations of traditional inspection methods—often inefficient, subjective, or reliant on costly offline equipment—a high-precision, real-time defect-detection model, YOLO-AMI, is proposed, based on the YOLOv10 architecture. The neck network was reconstructed using the Asymptotic Feature Pyramid Network to enhance multi-scale feature fusion and suppress background noise. In addition, a parameter-free attention mechanism was integrated to adaptively emphasize critical features without increasing computational complexity. To improve detection of small defects, a composite loss function combining Normalized Wasserstein Distance and Intersection over Union was adopted. Experimental evaluation on a dataset of 6,000 AM images shows that YOLO-AMI attains a mean average precision (mAP@0.5) of 85.5%, precision of 87.1%, and recall of 83.2%, outperforming state-of-the-art models such as YOLOv8, YOLOv11, and RT-DETR-L. With an inference speed of 105.6 frames per second and a compact model size of 8.6 million parameters, the proposed approach achieves a favorable balance between accuracy and efficiency, providing a robust solution for intelligent online quality monitoring in Industry 4.0.
Li et al. (Sat,) studied this question.