Semantic Segmentation of Steel Wire Rod Thermal Images for Automated Temperature Measurement

Hardness is a critical quality attribute of wire rods and is significantly influenced by cooling temperatures. To measure this temperature, infrared line scanners are typically employed to generate thermal images that provide full-length temperature data of wire rods. In these thermal images, some pixels correspond to the rod surface temperature, while others represent the background, which often contains noise. Accurately identifying rod-related pixels is essential, but this remains challenging due to the complex geometry of the wire rods. To address this, we propose a semantic segmentation method using a U-Net architecture to automatically distinguish wire rod pixels from the background. This method improves automation and eliminates the need for manual processing. We validated the approach on wire rods of varying diameters, from heavy-type (> 10 mm) to fine-type (≤ 10 mm). For heavy-type rods, our method achieved an accuracy of 0.9624 and a mean Intersection over Union of 0.8998. For fine-type wire rods, where smaller temperature differences complicate segmentation, we introduced a restoration technique using a sliding window approach, which enhanced segmentation quality. This combined method enables automated full-length temperature measurement across different wire rod types, supporting improved quality control in wire rod manufacturing.