Automated Detection and Localization of Steel Surface Defects Using a Fine-Tuned YOLOv8 Model

Automated surface inspection is an essential requirement in modern steel manufacturing because defects such as crazing, inclusions, patches, pitted surfaces, rolled-in scale, and scratches can reduce product quality, increase downstream rejection rates, and compromise process reliability. This study presents a deep learning-based steel surface defect detection framework using a fine-tuned YOLOv8m object detector on the Augmented NEU-DET dataset. The proposed approach formulates defect inspection as a multi-class object detection problem in which the model simultaneously localizes defect regions and predicts their corresponding categories. The dataset was organized in YOLO format and consisted of 2,852 training images and 363 validation images across six defect classes. Transfer learning was adopted by initializing YOLOv8m with pretrained weights, followed by supervised fine-tuning using an input resolution of 768 x 768 pixels, a batch size of 16, stochastic gradient descent optimization, an initial learning rate of 0.003, and early stopping patience of 80 epochs. Model performance was evaluated using precision, recall, mAP@0.50, mAP@0.50:0.95, learning-curve behavior, normalized confusion-matrix analysis, and qualitative validation-set inference. The trained detector achieved an overall precision of 0.803, recall of 0.719, mAP@0.50 of 0.800, and mAP@0.50:0.95 of 0.467. Class-wise results showed strong detection performance for scratches, patches, and pitted_surface, whereas crazing and rolled-in_scale remained more challenging due to their texture-like appearance and lower visual separability from the background. The results demonstrate that YOLOv8m provides a strong baseline for real-time steel defect localization and classification, while also indicating the need for further class-balanced training and localization refinement for visually ambiguous defect categories.