Two-step approach for fatigue crack detection in steel bridges using convolutional neural networks

The advent of parallel computing capabilities, further boosted through the exploitation of graphics processing units, has resulted in the surge of new, previously infeasible, algorithmic schemes for structural health monitoring (SHM) tasks, such as the use of convolutional neural networks (CNNs) for vision-based SHM. This work proposes a novel approach for crack recognition in digital images based on coupling of CNNs and suited image processing techniques. The proposed method is applied on a dataset comprising images of the welding joints of a long-span steel bridge, collected via high-resolution consumer-grade digital cameras. The studied dataset includes photos taken in sub-optimal light and exposure conditions, with several noise contamination sources such as handwriting scripts, varying material textures, and, in some cases, under presence of external objects. The reference pixels representing the cracks, together with the crack width and length, are available and used for training and validating the proposed model. Although the proposed framework requires some knowledge of the “damaged areas”, it alleviates the need for precise labeling of the cracks in the training dataset. Validation of the model by means of application on an unlabeled image set reveals promising results in terms of accuracy and robustness to noise sources.