Enhancing Crack Segmentation With Limited Data: SinGAN-Based Synthesis and Blending of Textures and Cracks

The reliable detection of structural cracks is crucial in maintaining the integrity of different structures like pipes and pavements, yet it is often constrained by the limited availability of diverse training datasets. This research presents a novel approach utilizing Single Image Generative Adversarial Networks (SinGAN) for the synthesis and blending of realistic textures and synthetic cracks, aimed at augmenting sparse datasets for improved crack detection. Recognizing the challenges in generating sufficiently varied and realistic data for training robust detection models, we leverage SinGAN’s capacity to generate textures from single images and introduce procedural techniques for the generation of synthetic cracks using Simplex noise. These synthetic cracks are seamlessly integrated with the generated textures, employing SinGAN’s harmonization capabilities, thus significantly enhancing the realism of our synthetic dataset. Another aspect of our research is the validation of the synthesized dataset’s utility in improving crack detection and segmentation. Utilizing a subset of “CrackSeg9k” dataset as a benchmark, we employed two state-of-the-art crack segmentation methods, Pix2pix and DeepLabv3+, to evaluate the efficacy of our approach. The inclusion of our SinGAN-generated data led to notable improvements in model performance in terms of F1 score and mean intersection-over-union (mIoU), emphasizing the value of our synthetic dataset in addressing the limitations posed by limited real-world data.