Crack detection in concrete bridges using deep learning
In Belgium, the majority of bridges was built over 60 years ago. The average lifespan of a bridge is around 70 years. This means that many bridges are reaching the end of their life. Bridge inspection is therefore crucial to detect signs of failure early on before a potential catastrophic failure. The process of inspecting a bridge is often difficult as they are hard to reach due to being over a motorway or train tracks. In this project, we validated the use of deep learning for automated crack detection using drones.
We worked together with domain experts from the Departement Mobiliteit en Openbare Werken (MOW) to learn the correct context and classification of cracks in bridges. Besides open available data, MOW provided us with a unlabelled dataset of their own. After labelling the dataset of MOW we used both datasets to train an image segmentation model. Image segmentation is a subbranch of computer vision which can determine per pixel if it belongs to a crack or not. This gives a fine-grained overlay of where the cracks are located.
The solution is able to detect cracks with a high accuracy in images of bridges. This project has provided MOW with the necessary insights to go ahead with the project. In the next phase we will look into more advanced techniques to classify the severity of the crack.