Steel surface defect detection and segmentation using deep neural networks
Sara Ashrafi, Sobhan Teymouri, Sepideh Etaati, Javad Khoramdel, Yasamin Borhani, Esmaeil Najafi
Abstract
Defect detection is a crucial task in the manufacturing industry, particularly in steel surface inspection. While manual recognition is one of the most reliable techniques, recent advances in computer vision and machine learning have led to the development of automatic defect detection techniques. This paper proposes several deep-learning-based computer vision techniques, including semantic segmentation and object detection models, to detect surface defects on steel sheets. The U-Net, FCN-8, and FPN models are implemented for segmentation, while the YOLOv4 model is used for object detection. Moreover, a combined segmentation and object detection structure, referred to as two-stage defect detection, is developed to enhance the accuracy of detecting small defects. Based on the obtained results, the U-Net model with pre-trained backbones achieves a Dice Similarity Coefficient of 72%, outperforming existing methods. The object detection model with a resolution of 640 reaches the mean average precision of 49.32% and 35.06% for binary class and multi-class detection, respectively. Furthermore, the proposed two-stage defect detection structure achieves a Dice Similarity Coefficient of 84%. In summary, the results validate the efficient performance of the studied techniques for accurate defect detection on steel surfaces. • Introduction of a robust framework combining U-Net, FCN-8, and FPN models for effective semantic segmentation of steel surface defects. • Extensive evaluation of object detection with YOLOv4 across multiple resolutions, highlighting the model's performance in binary and multi-class detection. • Addressing the challenge of imbalanced datasets by effectively training models to detect and segment various defect types, ensuring reliability in real-world applications. • Proposition of a novel two-stage defect detection method combining YOLOv4 for object detection and U-Net for segmentation, significantly improving small defect detection accuracy.