High quality training set development for road crack detection using progressive data-driven models with integrated identification

Generating convincing pavement crack imagery remains challenging due to the limited accessibility of annotated datasets and the uneven distribution of crack patterns in real-world conditions. Existing synthesis techniques often encounter unstable convergence and struggle to preserve fine structural textures. To address these challenges, this work proposes two adaptive image generation schemes that combine convolution-based feature extraction with a progressive refinement strategy. The first approach focuses on gradual spatial enhancement, while the second employs a dual-channel generation process that independently models pavement textures and crack structures before fusing them. Both methods were trained and validated on a standard crack image dataset, and their outputs were evaluated using perceptual and statistical indicators, including FID, KID, LPIPS, and SSIM. Results reveal that dual-channel design produces more coherent and realistic crack representations, improving visual fidelity and downstream detection accuracy compared to reference models. This study offers a scalable solution for augmenting limited road defect datasets, facilitating the development of stronger deep learning models for pavement condition monitoring under data-scarce environments.