Tatsuya　GOBARA

Study on the Application of Quantitative Deep Learning to Rust Evaluation of Weathering Steel Bridges

Eiji　IWASAKI

Weathering steel is a representative material of the corrosion protection methods utilizing corrosion-resistant materials. By adding appropriate amounts of alloying elements such as Cu, Ni, and Cr to ordinary steel, the steel surface forms a protective rust state under the appropriate corrosion environment. In order to confirm the formation of protective rust on weathering steel bridges, it is necessary to properly evaluate the state of rust, and rust evaluation, mainly through visual inspection, is often conducted. In recent years, quantification of rust evaluation using deep learning has been promoted in order to appropriately judge the rust condition, but there are still problems in classification based on feature extraction in images.
In this study, we investigated rust evaluation of weathering steel bridges based on rust thickness using deep learning. Specifically, we built the deep learning model by selecting a pre-trained model and optimizing hyperparameters, including the optimization algorithm and learning rate, and evaluated its classification performance. Additionally, we assessed the generalization performance of deep learning models for rust evaluation using images taken with smartphone cameras.
Through hyperparameter optimization, we found that a learning rate = 10E−6 was the most effective, regardless of the pre-trained model and optimization algorithm. The loss in training was found to be ViT-B/16 < VGG16 < ResNet50 for pre-trained models and Adam < RMSprop for optimization algorithms, indicating that ViT-B/16 and Adam performed the best. Also, the deep learning enabled rust evaluation, suggesting that maintenance can be made more efficient. In addition, Adjusting the RGB intensity of images suggests that rust evaluation is possible even when using different cameras or capturing images casually. By controlling the rotation around the Z-axis, stable classification can be expected within a rotation angle range of approximately 40°.