Data augmentation using CycleGAN-based methods for automatic bridge crack detection

Baoxian Li

doi:10.1016/j.istruc.2024.106321

2024-04-01

PRE

DOI：10.1016/j.istruc.2024.106321

原文链接

摘要

En 中文

Periodic crack detection is of great significance in preventing bridge failures and saving maintenance costs. In the past few decades, crack detection was highly depended on human-conducted on-site inspections. It was not transformed to automatic crack defection until a spot light on deep learning. Although automatic detection based on deep learning is of advantage in both of efficiency and accuracy, the heavy workload of manual annotation is the first challenge we face. Collecting rich and well-balanced crack defect datasets is the second one. To rise to the challenges, this paper proposes two improved CycleGANs, named Tiny Cycle-consistent Adversarial Network (Tiny-CycleGAN) and Multi Cycle-consistent Adversarial Network (Multi-CycleGAN). Additionally, inside of the neural networks we propose, we introduce spectral normalization into the discriminators of them to boost synthesis quality and stabilize training. Experimental results demonstrate that these approaches not only translate from bridge crack images to their paired high-quality label images, but also augment crack images from unpaired label images effectively. Lastly, a deep learning based semantic segmentation network, U-Net with attention gate (AttU-Net), is trained on the augmented dataset to evaluate the performance of improved CycleGANs. The evaluations of semantic segmentation show that dataset of augmentation with Multi-CycleGAN outperforms dataset without augmentation by up to 0.5417 on mIoU and 97.45% on mPA, respectively.

Keyword:

Bridge engineering

Deep learning

CycleGAN

Data augmentation

Crack detection