Robust Oriented Localization of Handwritten Identification Regions on Densely Stacked Steel Plates in Complex Industrial Scenes

Handwritten identification marks on steel plate end faces are important for material traceability, inventory checking, and production scheduling in large steel storage yards. Their localization remains challenging because field images often contain uneven illumination, specular reflection, rust, oil contamination, densely adjacent targets, and arbitrary target orientations. This study presents an application-oriented pipeline for oriented localization of handwritten marks on densely stacked steel plates. The pipeline combines task-oriented image enhancement, high-resolution sliding-window inference, and a YOLOv11 oriented bounding-box detector (YOLOv11-OBB) adapted with attention enhancement, adjusted multi-scale fusion, and Scylla-IoU (SIoU)-based box regression. A real steel-yard dataset was annotated with oriented bounding boxes and split at the source-image level before patch extraction. The held-out evaluation subset was used both for checkpoint selection and final performance reporting under the same protocol. On this dataset, the final detector achieved 89.9% precision, 88.3% recall, 91.8% mAP@0.5, and 72.0% mAP@0.5:0.95. These results support the practical value of oriented localization and high-resolution inference for weak handwritten regions in complex industrial scenes. Broader detector benchmarking and downstream recognition performance remain to be validated in future work.