Crack identification for bridge condition monitoring using deep convolutional networks trained with a feedback-update strategy

Orthotropic steel bridge decks and steel box girders are key structures of long-span bridges. Fatigue cracks often occur in these structures due to coupled factors of initial material flaws and dynamic vehicle loads, which drives the need for automating crack identification for bridge condition monitoring. With the use of unmanned aerial vehicle (UAV), the acquirement of bridge surface pictures is convenient, which facilitates the development of vision-based bridge condition monitoring. In this study, a combination of convolutional neural network (CNN) with fully convolutional network (FCN) is designed for crack identification and bridge condition monitoring. Firstly, 120 images are cropped into small patches to create a basic dataset. Subsequently, CNN and FCN models are trained for patch classification and pixel-level crack segmentation, respectively. In patch classification, some non-crack patches that contain complicated disturbance information, such as handwriting and shadow, are often mistakenly identified as cracks by directly using the CNN model. To address this problem, we propose a feedback-update strategy for CNN training, in which mistaken classification results of non-crack data are selected to update the training set to generate a new CNN model. By that analogy, several different CNN models are obtained and the accuracy of patch classification could be improved by using all models together. Finally, 80 test images are processed by the feedback-update CNN models and FCN model with a sliding window technique to generate crack identification results. Intersection over union (IoU) is calculated as an index to quantificationally evaluate the accuracy of the proposed method.