Generation of Infrastructure Crack Images for Self-Supervision Training Based on Diffusion Model

Data scarcity often hinders the application of artificial intelligence techniques for detecting defects in infrastructure. To address this issue, a pure crack image generation method is proposed based on the diffusion algorithm. This method can produce images with various crack features in real backgrounds and improve the accuracy of the detection model by enhancing the training data. Compared to generative adversarial networks (GAN), the proposed method does not require a discriminator or labeled samples for training, thus generating more diverse results. To improve the sensitivity of the generative model to crack features, a specialized denoising model is developed that simultaneously employs convolution and transformer for feature extraction branching. These two types of feature maps are effectively fused by proposing a deformable feature extraction block (DFEB) and self-selecting feature fusion block (SFFB) so that the cracks and the backgrounds receive sufficient attention at the same time. A staged training strategy is used to merge the crack and background styles of multiple datasets, which further increases the diversity of the model. The experimental results show that the Fréchet inception distance (FID), generation precision (GP), generation recall (GR) and generation F1 score (GF1) of the proposed crack generation method are 3.10, 0.72, 0.75 and 0.73, respectively, which is more advantageous than the existing diffusion models and GAN models. In addition, an existing defect segmentation model is tested using the generated crack images based on self-supervised transfer learning. This model improves the accuracy (Acc), F1-score (F1), and intersection over union (IoU) of the target dataset by 9.18 percentage points (pp), 6.99 pp, and 7.83 pp, respectively. The experimental results demonstrate the significant effect of the proposed method on improving the detection accuracy of small-scale datasets. The trained model has been made into a crack image generation tool that is open-sourced for validation and use by other scholars.