Design, implementation, and deployment of a machine vision-based dimension measurement system for steel slabs using deep neural networks

Accurate measurement of steel slab dimensions is important for quality control and process monitoring in hot rolling. This study presents an automated machine vision system for measuring slab length, width, and thickness under real production conditions. The system combines a dual-camera imaging setup, controlled illumination, DeepLabV3 + semantic segmentation with an EfficientNet-B5 backbone, RANSAC-based boundary fitting, and a height-dependent calibration strategy for pixel-to-millimeter conversion. The software is integrated with the plant’s SQL-based network to compare measured dimensions with Product Data Information (PDI) and provide operator feedback. The system was evaluated on 65 previously unseen steel slabs that were independent of the 1,514 slabs used for network training and validation. Mean absolute errors were 2.56 mm, 1.84 mm, and 1.04 mm for length, width, and thickness, respectively, with corresponding RMSE values of 3.15 mm, 2.39 mm, and 1.35 mm. Using plant-specific tolerance bands of ± 5 mm, ± 3 mm, and ± 2 mm, within-tolerance agreement rates were 98.5%, 96.9%, and 95.4%. The complete measurement cycle takes approximately 4.2 s per slab and the system is currently deployed on a production line.