Minimizing waste via novel fuzzy hybrid stacked ensemble of vision transformers and CNNs to detect defects in metal surfaces

Surface defect detection in steel manufacturing is crucial for quality control and waste reduction, yet traditional manual inspection methods are time-consuming, subjective, and increasingly inadequate for modern production demands. This paper introduces a novel fuzzy hybrid stacked ensemble model that combines vision transformers (ViTs) and convolutional neural networks (CNNs) to detect and classify surface defects in hot-rolled steel strips. Our approach integrates multiple advanced preprocessing techniques including Canny Edge detection, Gaussian Blur, cumulative distribution function (CDF), and local binary patterns (LBP) with state-of-the-art deep learning architectures. The ensemble leverages a Gompertz function-based fuzzy ranking system to optimally combine predictions from individual models, addressing common challenges such as inter-class similarity, intra-class variance, and imbalanced datasets. We evaluate our model on the NEU-CLS-64 dataset, which contains over 7000 grayscale images across nine defect categories. The proposed ensemble achieves exceptional performance with 99.45% accuracy, significantly outperforming existing methods, including individual ViTs (95.68%) and CNNs (92.96%). Our model demonstrates robust generalization capabilities and maintains high precision (96.33%) and recall (94.16%) across all defect classes, making it particularly effective for real-world industrial applications. This research contributes to the growing field of automated quality control in steel manufacturing, offering a reliable solution that aligns with Industry 4.0 principles while addressing the limitations of traditional inspection methods.