Image-driven bridge inspection framework using deep learning and image registration

This paper proposes an image-driven bridge inspection framework using automated damage detection using deep learning technique and image registration. A state-of-the-art deep learning model, Cascade Mask R-CNN (Mask and Region-based Convolutional Neural Networks) is trained for detection of cracks, which is a representative damage type of bridges, from the images taken from a bridge. The model is trained with more than a thousand training images containing cracks as well as crack-like objects (hard negative samples). The images taken from a test bridge are input to a deep learning model trained to detect damages, which is further mapped on a large image of each bridge component registered using a commercial registration software. The performance of the proposed framework is evaluated on piers of existing bridges, whose external appearance was imaged using a DSLR with a telescopic lens. The results are compared with the conventional visual inspection to analyse the performance and applicability of the proposed framework.