An attention-guided dual-branch global–local deep learning framework for multiclass surface defect classification in steel billets

Automated inspection of surface defects in steel billet cross-sections is challenging in industrial production lines because of high processing speeds, illumination variations, thermal reflections, and imbalanced defect classes. This study proposes a Grad-CAM-guided dual-branch Global–Local deep learning framework for multiclass billet defect classification. In the proposed architecture, a global branch processes the entire image to extract structural and contextual representations, whereas Grad-CAM attention maps generated from a target convolutional layer automatically localize the most discriminative region of interest. The localized region is then analyzed by a local branch to capture fine-grained texture patterns and subtle defect-related features. Global and local representations are fused through feature concatenation and fed to a classification head for final prediction. The dataset consists of real billet cross-section images acquired from the production line of Khuzestan Steel Company and includes seven categories: six defect types and one healthy class. An embedded hardware platform was also developed to enable real-time deployment and automatic marking of defective billets under harsh industrial conditions. Robustness and generalization were evaluated using stratified 5-fold cross-validation with MobileNetV2, ResNet50, and DenseNet121 backbones. ResNet50 achieved the best performance, with 93.05% accuracy and a macro AUC of 0.9888, while MobileNetV2 remained suitable for resource-constrained deployment.