Scale–Space Data Augmentation for Deep Transfer Learning of Crack Damage from Small Sized Datasets

Abstract

Crack damage is commonly observed for civil structures and infrastructure in service. The recent years have witnessed an excessive utilization of deep learning models for realizing autonomous and machine-vision based crack detection. Given this trend, this paper recognizes two entwined challenges: the preparation of large-scale training data and the detection of simple crack damage amid complex scenes. To address them, a novel data augmentation technique is proposed considering crack characteristics in images for realizing deep transfer learning using very small datasets. Numerical experimentation is conducted based on two types of crack datasets (concrete structures and asphalt pavement), each of which has only tens of images containing complex scenes. When evaluating the performance, a sliding-window based rating scheme is proposed, which is much stricter than the conventional bounding-box based approach. Quantitative performance analysis shows the acceptable performance (e.g., an overall accuracy of 93.81%, an F-2 score of 74.4%, and a very high recall of 91% for the crack detection in concrete images). The result demonstrates the effectiveness of the proposed data augmentation method and the superior transferability if the transfer learning is carried out through a fully fine-tuned training process.