Corrosion of roadway metal railings in seaside cities not only affects the cosmetic appearance but also
brings an economic cost to local government. This study aims to inspect and evaluate the corrosion of
roadside guardrails with deep learning technique. Hundreds of photos on corroded roadway metal
railings are taken, considering various illumination conditions, diverse shooting angles and different
shooting distances in order to improve the generalizing ability of the deep learning model. The pictures
are annotated manually labelling damages pixel by pixel, and DeepLabV3+ is trained to semantically
segment the pictures for detection of corrosion shape and location. A subset of the photo dataset which
has not been used in the training process is tested for validation, and two models with different loss
function were trained to compare the influence of the adopted loss function. Based on the validated
photos, the DeepLabV3+ model can detect corrosion area present in the image, which is useful to help
the inspectors to design the maintenance strategy for the roadway metal railings.