This study deals with the comparison of different versions of YOLO and then addressing the problem of detecting surface defects in load-carrying rails within automobile assembly environments, which are industrial environments requiring very high precision and reliability. In order to address the challenges posed by the complexity and variability of such defects, we introduce a novel detection system based on an enhanced YOLOv5 architecture specifically developed to meet the rail surface inspection needs. The detection system utilizes advanced computer vision strategies in conjunction with deep learning enhancements to provide accurate and robust detection of defects. The Multi-Scale Pyramid Pooling (MSPP) module is a significant part of the system that extracts rich features at multi-scales, using residual stacking. Additionally, we introduced a new Dual Attention Mechanism (DAM) to re-focus the model's attention on small and irregularly shaped surface defects. This has been one of the challenges with traditional models.
Experimental results validate the adequacy of the framework, as evidenced by the adjacent model’s AP50 of 97.3%, representing a 4.2% improvement over the former YOLOv5. The adjacent model has a mean Average Precision (MAP) of 88.9%, with a corresponding recall of 91.4%, and an accuracy of 92.6%, all demonstrating competent performance across important parameters. The results emphasize the possibility of the framework's practical use in automatic rail defect detection; as a reliable solution for quality assurance in an industrial domain, the framework is shown to be highly accurate.