Intelligent detection methods for miniature defects in metallic materials

Powder metallurgy (PM) technology is extensively used in high-value industries for its energy efficiency, precision, and cost-effectiveness. However, detecting mini-defects in PM production remains challenging, particularly in highly customized and small-batch productions. This study evaluated defect detection in PM parts, PMPDv1 and PMPDv2 datasets comprising 457 and 1521 images, respectively. Automated optical inspection (AOI) and image augmentation techniques were applied to enhance image quality and model learning. The YOLO series models were employed for automated defect detection.

Results demonstrated that YOLOv4 achieved a mean average precision (mAP) of 93.94% at a resolution of 1600 but required 31 GB of GPU memory and 881,443 GFLOPs. YOLOv5s, under the same conditions, achieved an mAP of 92.7% with just 12.1 GB of GPU memory and 15.8 GFLOPs, making it suitable for resource-constrained environments. This study confirms the efficacy of YOLO models for PM defect detection and suggests further exploration of transfer learning and generative AI techniques to enhance detection efficiency and accuracy in other products.