A data augmentation method for pavement crack detection based on super‐resolution and denoising diffusion probabilistic models

Automated detection of pavement cracks is a task of wide interest. With the improvement of industrialization, high-resolution (HR) images are increasingly favored by researchers due to their ability to provide rich information about pavements and diseases. However, the acquisition of effective training data is not easy, which affects the accuracy and robustness of the detection model. Although the recently emerged denoising diffusion probabilistic model (DDPM) overcomes the inherent pattern collapse problem of generative adversarial networks and is capable of generating more diverse and realistic pavement data, its high sampling cost hinders the generation of HR images with rich texture information. To overcome this limitation, this paper proposed a low-cost, two-step data augmentation method that combines DDPM with super-resolution. The method first generated small-sized pavement crack images using DDPM and then enhanced resolution and texture details using an improved SwinIR model. The resulting HR and diverse crack images were used to augment the dataset. The effectiveness of the proposed method was evaluated using four state-of-the-art object detection models. Experimental results showed that all models trained with the augmented training dataset exhibited better performance. Furthermore, when combined with geometric transformation techniques, the proposed method was able to improve the crack detection accuracy by up to approximately 12%.