Enhancing Structural Cracks Segmentation Through Generative Image Synthesis

Using deep neural networks to detect cracks in architectural structures presents significant challenges due to the irregular and diverse nature of cracks, along with variations in lighting conditions and backgrounds. Moreover, the absence of a sufficient number of crack sample images can significantly impact the accuracy of crack detection. In this study, we propose a novel method to enhance the semantic segmentation accuracy of structural crack images. This is achieved by generating images and their corresponding masks. Our approach cleverly combines synthetic crack data generated through unconditional GAN-based methods with real-world data, addressing the issue of limited data availability. Additionally, we employ transformer-based architecture, specifically the SegFormer segmentation model, to assess our methodology. Experimental results demonstrate that the SegFormer small model, trained iteratively on this integrated dataset, achieved an impressive mIoU of 87.13%, marking a significant improvement over the 85.59% obtained on the original dataset. Our work demonstrates the power of generative techniques in computer vision tasks by offering a viable solution in the face of challenging data constraints, especially in the realm of cracks detection.