Pavement Crack Segmentation Based on Synthetic Data Sets and Unsupervised Domain Adaptation

This study proposes a novel pavement crack segmentation methodology that integrates synthetic data sets with unsupervised domain adaptation to address annotation challenges in pavement crack data sets and enhance segmentation performance. An automated program has been developed to generate synthetic pavement crack images along with high-precision semantic segmentation labels at a speed of 0.5 s per image. These synthetic images accurately replicate the irregular geometry, randomness, and variability of real pavement cracks. The CycleGAN unsupervised domain adaptation network, which incorporates the Structural Similarity Index Measure (SSIM) in its loss function, ensures that the translated images from real crack images preserve the structural characteristics of real images while enhancing domain adaptability. Applying a segmentation network to the translated images effectively mirrors segmenting real images. Segmentation performance across six semantic segmentation networks—U-net, Erfnet, DeepLabV3, FCN, PSPNet, and SegFormer—trained with synthetic images shows significant improvements when applied to the translated images, compared with directly segmenting real images, with average increases of 16.55% in mFscore and 12.62% in mIoU. Moreover, 10 domain adaptive segmentation networks—Daformer, Hrda, IAST, CBST, PyCDA, TransNorm, CLAN, FADA, AdaptSeg, and DDC—are evaluated to assess the performance of the synthetic data set. Among these, Daformer exhibits exceptional performance with an mFscore of 79.52% and an mIoU of 69.67%. In addition to the CRACK500 data set, the generalizability of the synthetic data set is validated using two other real pavement crack data sets: AEL and GAPs384. An ablation study on the CycleGAN network, trained both with and without the synthetic data set, demonstrates that incorporating the synthetic data set significantly enhances the quality of the translated images, thereby improving the model’s segmentation performance on real crack data sets. These findings highlight the significant potential of synthetic data sets in semantic segmentation tasks.