Abstract: Deep learning-based methods have demonstrated a remarkable performance in the segmentation of cracks occurring in concrete structures. However, their effectiveness remains constrained by the quantity and quality of training datasets. To address the high acquisition cost and labor-intensive nature of existing crack datasets, this study proposes a lightweight crack image synthesis method, which utilizes a stochastic fractal noise algorithm to generate foreground crack patterns, and introduces a foreground-background fusion strategy based on the radial distance from the crack axis. The method proposed enables the efficient and automated generation of high-quality crack images along with their corresponding annotations. To validate the method proposed, neural network models were trained under identical configurations using real data, synthetic data, and a combination of both, and tested on a real-world crack dataset. The simulation results show that, under the same data volume, the performance gap between models trained on synthetic data and real data remained within 10% across all key metrics. When using a hybrid dataset, the discrepancy in metrics such as Intersection over Union (IoU) decreased to approximately 2%, and the recall rate slightly surpassed the real-data model by 0.33%. Furthermore, the average generation time per synthetic image was only 136 milliseconds, outperforming conventional manual annotation methods. This study offers a cost-effective and highly efficient solution for generating training data for image-based structural crack segmentation tasks, especially in engineering scenarios where real data is limited.