Steel surface defect detection is a core component of quality control in steel manufacturing and is crucial for product performance and production efficiency. In recent years, computer-vision-based deep learning methods have decisively surpassed traditional machine-vision approaches built on handcrafted features, propelling steel defect detection into the intelligent era. The review itself is limited to deep learning methods; handcrafted machine vision is background only. This paper provides a comprehensive survey of deep-learning-based steel surface defect detection, citing 184 publications, of which 183 were published between 2015 and 2026. We organize the literature around the three core tasks—classification, detection, and segmentation—and review each task from the perspective of network architecture optimization. We further summarize more than ten public datasets, including NEU and GC10-DET, together with the commonly used evaluation metrics and open-source code resources, and compile literature-reported results for dozens of models, including the YOLO series, DETR, and Swin Transformer. Those figures are copied from the original papers and are not a unified benchmark. Finally, we analyze the core challenges—data scarcity, small-target detection, model generalization, and real-time performance—and discuss future research directions, including high-resolution image processing, few-shot learning, and large-model-assisted quality assessment.