Innovative synthetic data augmentation for dam crack detection, segmentation, and quantification

Although deep-learning-based approaches have demonstrated impressive performance in object detection tasks, the requirement for large datasets of annotated training images limits the feasibility of deep neural networks. For example, obtaining a large number of crack images of a dam is unlikely, particularly in the absence of open-source datasets. To address this problem, the authors have developed three synthetic data generators based on virtual scene simulation and image processing for generating large amounts of labeled dam surface crack data. These synthetic data combined with public-available images of cracks on pavement and concrete are further used to train a state-of-the-art object detection neural network, resulting in a 29.2% improvement in the overall crack detection mean average precision (mAP) compared to using only images of cracks on pavement and concrete. Furthermore, given the necessity for further analysis of some critical cracks, an image-processing-based approach for segmenting the crack in each detected bounding box and estimating its length and thickness is provided.