Comparison of Deep Learning Methods and a Transfer-Learning Semi-Supervised GAN Combined Framework for Pavement Crack Image Identification

Abstract

The pavement crack identification performance of typical models or algorithms of transfer learning (TL), encoder-decoder (ED), and generative adversarial networks (GAN), were evaluated and compared on SDNET2018 and CFD. TL mainly takes advantage of fine-tuning the architecture-optimized backbones pre-trained on large-scale data sets to achieve good classification accuracy. ED-based algorithms can take into account the fact that crack edges, patterns or texture features contribute differently to the identification. Both TL and ED rely on accurate crack ground truth (GT) annotation. GAN is compatible with other neural network architectures, thus can integrate various frameworks (e.g., TL, ED), and algorithms, but the training time is longer. In patch classification, the fine-tuned TL models can be equivalent to or even slightly better than the ED-based algorithms, and the predicting time is faster; In accurate crack location, both ED- and GAN-based algorithms can achieve pixel-level segmentation. It is expected to realize real-time automatic crack identification on a low computational power platform. Furthermore, a weakly supervised learning framework (namely, TL-SSGAN) is proposed, combining TL and semi-supervised GAN. It only needs approximately 10%–20% labeled samples of the total to achieve comparable crack classification performance to or even outperform supervised learning methods, via fine-tuned backbones and utilizing extra unlabeled samples.