BRIDGE CORROSION DAMAGE DETECTION USING DEEPLABV3+ MODEL AND PERFORMANCE BOOSTING

Abstract

In recent years, the deterioration of bridges has become a major problem in many countries. Specially in Japan, the infrastructures are required to be inspected at least once per five years, so that they can to be maintained properly. However, the number of civil engineers is continuously decreasing, and conventional inspection methods require a lot of skilled manpower, equipment, and time. Nowadays, an improved approach has been proposed to use UAV for data gathering and different AI techniques for bridge damage detection. The AI detection of multiple damages is not sufficiently accurate, and the accuracy decreases when the number of damage types increases. In this study, DeepLabv3+ model was used to detect a single damage that focused on corrosion. Transfer learning was used to improve the accuracy of the boundaries. The effect of annotation's quality on the trained model, improvement of accuracy by finding the optimum hyperparameters, and the reduction of false positives by adding background photos were discussed in this study.