Controllable reference-based semantic crack-image generation using diffusion model for intelligent infrastructure inspection

Abstract

Improving the accuracy and robustness of deep-learning-based crack-segmentation models remains a significant challenge, primarily because of the insufficient quantity and diversity of the available pixel-level annotated data. To address this issue, this paper proposes a controllable Crack Reference-based Diffusion Model (CRDM). The proposed model can accurately synthesize realistic cracks on crack-free background images by leveraging predefined masks and reference images. Notably, it effectively transfers crack features from reference images to generated images, while maintaining high semantic accuracy. Extensive experiments are performed to demonstrate the advantages of CRDM in producing high-quality, diverse, crack images with precise controllability. The dataset augmented with the CRDM-generated images improves the performance of crack-segmentation models by ∼1 % IoU, across various scenarios. Further performance gains are achieved through our refined label-filtering strategy. The proposed CRDM exhibits strong potential for crack-segmentation tasks, effectively reducing the time and cost of data annotation and acquisition.