Character recognition method for angle steel based on the optimized YOLOX framework

Abstract

The character recognition of angle steel in factory environments is often challenged by factors such as noise, rust, and varying lighting conditions. To address this, we explore the application of the YOLOX model and propose an enhanced version, T-YOLOX. By integrating the Ghost module and FReLU activation, T-YOLOX achieves model lightweighting and improves the ability to capture character features, thereby enhancing recognition speed and accuracy. The effectiveness of these improvements was validated through ablation experiments. Moreover, we conducted comparative experiments between T-YOLOX, several YOLO models, and other mainstream object detection algorithms, demonstrating that T-YOLOX performs better across multiple key evaluation metrics. Furthermore, we developed two practical applications: an embedded recognition system for production lines and a mobile handheld system for real-time recognition in outdoor and distributed scenarios, both supporting seamless MES integration. These results extend the application of YOLOX in industrial settings and provide robust technical support for angle steel character recognition in steel tower manufacturing.