An Edge Is All You Need: Cracking the Code for Generating Synthetic Datasets for Robust Crack Detection Models

Generating synthetic datasets for training crack detection models remains a challenge due to the variability of crack appearances and the need for pixel-level annotations. As a promising alternative to manual labeling, synthetic data generation using Perlin Noise has gained attention. However, the relationship between Perlin Noise parameters and the physical characteristics of cracks is not well-established, leading to potential dataset bias and suboptimal model performance. This paper presents a novel approach that maps Perlin Noise generated crack parameters to crack characteristics such as width, length, curvature, and bifurcations, enabling the generation of diverse and realistic crack patterns. We employ a broad domain randomization technique by projecting the generated cracks onto randomly selected background images from the ImageNet database to reduce dataset bias and enhance model robustness. Using the Partitioning Around Medoids (PAM) algorithm, we create six distinct datasets capturing a comprehensive range of crack parameter variations. We demonstrate the effectiveness of our approach by fine-tuning the state-of-the-art SegFormer 5b model on our synthetic datasets and benchmarking its performance on the Crack500 dataset. Through linear regression analysis, we can identify the key crack parameters that influence model performance. Our results show that synthetic cracks have to be long, wide, straight, and should have few bifurcations to maximize evaluation metrics such as Intersection over Union (IoU), precision, recall, and F1.