Performance evaluation of CNN models for steel surface defect detection in lean manufacturing

The assurance of high-quality steel is critical in industries such as construction, automotive, and machinery. Surface defects in steel products can compromise product integrity, leading to waste and inefficiencies that are antithetical to Lean manufacturing principles. Traditional manual inspection methods are inefficient and subjective, hindering production efficiency. Automated inspection using computer vision techniques, particularly Convolutional Neural Networks (CNNs), has shown significant promise in accurately identifying defects, thereby supporting Lean initiatives by reducing waste and improving process efficiency. This study conducts a comprehensive analysis of various CNN-based image classification algorithms—including MobileNet, ShuffleNet, DenseNet, RegNet, NasNet, ResNet, GoogleNet, and a custom-designed CNN model—for classifying steel surface defects using the NEU-CLS-64 dataset. Performance is evaluated using key metrics such as accuracy, precision, sensitivity, and F1 score. Our results highlight the outstanding performance of RegNet, achieving the highest accuracy, precision, sensitivity, and F1 score. However, this elevated performance comes at a higher computational cost than other models, with MobileNet showing the least compelling performance. By integrating advanced defect detection into Lean manufacturing processes, this study demonstrates how selecting the appropriate CNN model can enhance production efficiency, minimize waste, and contribute to continuous improvement in metal production.