A parameter-free and effective automatic augmentation method for improved crack segmentation

Deep learning methods, such as semantic segmentation, have shown excellent performance and promising application prospects for detecting cracks in infrastructure. However, the substantial costs associated with the acquisition and expansion of labeled training datasets have significantly constrained model performance. Data augmentation is an ideal method to mitigate this problem. However, the transferability of specific augmentation operations is poor, and the choice of augmentation strategies often requires trade off complexity, cost, and performance. In this paper, a new region-based augmentation operation Mix-Crack and a novel augmentation selection method based on liner regression are proposed. Furthermore, with the support of the previous two methods, a simple, effective, and general automatic augmentation method is presented for the crack segmentation, named HybridAugment (HA). The entire process of HA is parameter-free and cost-effective. Extensive experiments were conducted to study its performance. First, thorough ablation experiments were conducted with different operational designs, augmentation space sizes, and sample compositions. And then, the HA was compared with previous state-of-the-art methods in a variety of crack segmentation scenarios. The results demonstrate that the HA outperforms previous augmentation methods, and achieves an excellent performance in the accuracy of crack prediction and the stability of model training.