Enhancing Steel Manufacturing Quality Control: A Deep Learning Approach for Rapid and Accurate Surface Defect Detection Using Pre-Trained Convolutional Neural Network Features and Supervised Classifiers

Abstract

The importance of perfect surface quality in flat steel manufacturing has grown; however, traditional computer vision approaches for automatic fault spotting are inaccurate and slow, lagging behind assembly line needs. This research investigates using pre-trained Convolutional Neural Networks (CNNs) as feature extractors to reliably categorize steel surface flaws. Sample images containing defects in hot-rolled steel strips were obtained from JSW Steel Ltd’s Vijayanagara Works, Ballari, Karnataka, India. Features were extracted from the penultimate layers of VGG16 and DenseNet-201 models before feeding the data to supervised classifiers (Support Vector Machine, Multilayer Perceptron, Random Forest) to categorize defects. We examined performance across model configurations based on combining the pre-trained CNNs and classifiers. The DenseNet-201+MLP model achieved maximum accuracy of 98.31%, outperforming VGG16 combinations and highlighting the promise of this approach. Thorough statistical evaluation via F-1 scores, precision, recall, etc. further demonstrates the value of using pre-trained CNNs with classifiers for manufacturing quality control, significantly improving steel surface defect detection speed and accuracy over traditional computer vision. This strategy shows promise for enabling reliable, large-scale automated inspection critical for industrial steel production.