<p>Reliable detection and classification of surface defects in hot-rolled steel strips are essential for ensuring product quality and reducing inspection costs in modern manufacturing. This study presents a comprehensive deep learning framework that integrates convolutional, transformer, and one-stage detection models for robust defect recognition. The Northeastern University (NEU) dataset, containing six representative defect types, served as the benchmark. For image-level classification, Xception and Vision Transformer (ViT) models were trained using extensive preprocessing, transfer learning, and interpretability analysis. Both achieved 100% test accuracy, with ViT converging ~ 4.5× faster and offering ~ 3× lower inference latency than Xception. For defect localization, a YOLOv8n model trained end-to-end achieved mAP50 = 72.4% and mAP50–95 = 44.2% at 42.7 ms/image, demonstrating strong suitability for real-time deployment despite the higher complexity of localization tasks. Comparative results show that while CNNs and transformers provide near-perfect classification precision, YOLOv8n offers the best balance between detection accuracy and speed, making it well-suited for industrial inspection lines. Grad-CAM explainability further confirmed that the models consistently focused on the true morphological regions of each defect type, enhancing transparency and strengthening trust in real-world applications. Overall, this work demonstrates that hybrid deep learning pipelines combining classification and detection can deliver both high accuracy and operational scalability for intelligent steel surface inspection, supporting the broader adoption of digital manufacturing solutions in Industry 4.0.</p>