2026 Volume 7 Issue 3 Pages 140-148
This study investigates the applicability of a method for estimating corrosion-induced section loss from internal surface images of underground steel conduits. Steel conduits removed from actual sites were used as specimens. Internal surface images showing corrosion were acquired using a scope camera, and the corresponding corroded locations were cut out as specimens. The cross sections of the cut specimens were then measured using scanning electron microscopy to obtain measured section loss values. A dataset of 437 pairs of internal surface images and measured section loss values was constructed. An ensemble learning approach based on MobileNetV2 image regression models was applied to estimate section loss from the internal surface images. For 30 test samples, the proposed method achieved a correlation coefficient of 0.803, a mean absolute error of 0.202 mm, and a root mean squared error of 0.235 mm. Within the scope of this validation, the results indicate that corrosion-induced section loss in underground steel conduits can be estimated from internal surface images at the submillimeter scale and that the estimates may serve as supplementary information to complement conventional visual assessments.