Defect Detection and Classification on Hot-Rolled Steel Using Machine Learning and Deep Learning Algorithms: A Comparative Analysis

Abstract Elevated concerns regarding sustainable manufacturing have resulted in increased efforts to deploy data-driven methodologies incorporating automated systems for fault analysis. In particular, manufacturing is increasingly focused on creating systems that detect and categorize defects, facilitating root-cause investigations. This research paper delves into the use of machine learning (ML) and deep learning (DL) approaches for defect detection in hot-rolled steel, focusing on examining the robustness of the different defect detection techniques. In the case of ML approaches, three primary feature extraction techniques: local binary pattern (LBP), histogram of oriented gradients (HOG), and gray-level co-occurrence matrix (GLCM), were employed. Generally, GLCM demonstrated relatively good performance, i.e., attained precision, recall, and f1-score values of 0.91 with a support vector machine (SVM) classifier. Similarly, using SVM, LBP attains precision, recall, and f1-score values of 0.89. Deep-learning methodologies such as convolutional neural networks (CNN), CNN plus VGG19 (CNN + VGG19), and you only look once version 7 (YOLOv7) were employed to investigate and classify the hot-rolled steel defects. CNN + VGG-19 and YOLOv7 exhibited excellent defect detection performance, attaining accuracy values of 0.9639 and 0.915, respectively. Overall, the results demonstrated relatively higher performance can be derived from using deep learning approaches such as CNN + VGG19 compared to traditional machine learning approaches.