AI spots steel defects at factory speed, matching human eyes
Researchers have developed a lightweight AI system that identifies surface defects on steel sheets moving at 100+ meters per minute—impossible for human inspectors. The technology achieves 99.7% accuracy while running on resource-constrained hardware, enabling real-time quality control at scale without slowing production.
Originaltitel: Defect attention-based lightweight real-time adaptable surface defect analysis
<p>Surface defect diagnosis for steel sheets is essential to maintain the quality of the product. However, it is challenging to do such a task due to the high manufacturing speed of 100-140 m/min, making it impossible to observe defects visually. However, developing a lightweight and real-time adaptable intelligent approach can both overcome this issue and satisfy the requirements of the concept of Industry 4.0. Based on that motivation, this study aimed to satisfy such a demand by introducing an attention-based surface defect identification approach called MobileNetV2-AGCA and UNET-AGCA for defect localisation. The effectiveness of the proposed techniques was evaluated by considering steel strip surface defect images of the NEU-CLS and NEU-DET surface defect dataset where six different defects are included. The Adaptive Graph Channel Attention (AGCA) Module was implemented in an edited version of MobileNetV2 and UNET architecture to improve segmentation accuracy while maintaining or improving the execution speed. According to the experimental results, MobileNetV2-AGCA successfully identified the surface defects with an average accuracy of 99.72%. Regarding defect localisation, UNET-AGCA reached an F1 score of 0.795. It is concluded that the proposed methods provided promising results to be applied to a real-time problem related to surface defect identification and localisation.</p>