Optimizing deep learning models for fast and effective rust identification

Steel bridges are highly susceptible to corrosion, which threatens structural safety and requires reliable inspection methods. Traditional manual inspections lack accuracy, consistency, and real-time capability. This study proposes an efficient rust detection framework that integrates HSV color space transformation, a U-Net (ResNet-34) deep learning model, and a webcam. To address the challenge of annotating dispersed rust spots, HSV-based auxiliary labeling was applied, with manual supplementation for cases involving light/dark rust colors. Experimental results demonstrate that the proposed method significantly improves labeling efficiency while maintaining high accuracy. The model achieved IoU of 0.69–0.87, and F-measure of 0.93–0.99 across different scenarios, outperforming approaches such as LS-SVM, ANLDR, and ANNRI. Integrated with a webcam, the system completed rust detection in approximately 7.5 seconds per image, enabling real-time application. The findings confirm that even with a relatively small dataset, U-Net (ResNet-34) provides fast, accurate, and stable rust detection, highlighting its potential for practical steel bridge monitoring and maintenance.