2026 Volume 7 Issue 3 Pages 831-843
There is a growing demand for improved efficiency in the inspection of road appurtenances―such as road signs that provide traffic information and lighting installed to illuminate roads―due to their enormous number and vast installation area. While previous models classifying the type and severity of damage to road appurtenances have achieved a certain level of accuracy, they have faced issues such as the risk of overfitting due to training on small-scale datasets, task settings limited solely to classification, and a lack of empirical analysis. Therefore, in this study, we construct a representation learning technique from images of road appurtenances trained on a medium-scale dataset. After proposing technologies for damage classification and the presentation of similar cases, we report the results of an empirical analysis conducted in FY2025 (Reiwa 7). Validation using a portion of the medium-scale dataset as validation data suggested that the proposed method can classify damage types and severity with high accuracy. Although empirical validation also suggested a certain degree of effectiveness, a degradation in accuracy due to factors such as domain shift was confirmed. On the other hand, feedback from field workers indicated the practical utility of the model’s attention regions and its ability to present similar cases.