Weakly-Supervised Pavement Surface Crack Segmentation Based on Dual Separation and Domain Generalization

Automatic pavement surface crack segmentation is crucial for efficient and cost-effective road maintenance. Despite fully-supervised crack segmentation methods have achieved significant success, the laborious task of pixel-level annotation hampers their widespread applicability. To address this issue, this paper presents a weakly-supervised pavement surface crack segmentation method based on Dual Separation and Domain Generalization (DSDGNet). Firstly, a crack image formulation model (CIFM) is developed by separating the crack image into a background component and a crack component. Additionally, we treat the crack component as a linear fusion of the pavement texture component and the crack mask. Secondly, a local-to-global learning method (L2G-L) is proposed to learn complete crack via local learning based on a random cropping and pasting algorithm. This idea stems from the observation that the crack component can be separated into several local regions, akin to the local regions found in hand-drawn crack components. Thirdly, A progressive interaction training algorithm (PIT) is crafted to train the image generation model by leveraging both generated and real images, thereby narrowing the divide between generated and authentic crack images. Finally, realistic and diverse crack images, along with their crack masks, are generated to facilitate the training of fully-supervised segmentation models. A generalizable loss is proposed to enhance the model generalization ability by combining reconstruction, segmentation, and domain adversarial losses. Extensive experiments on six public pavement crack datasets show the effectiveness and superiority of DSDGNet in weakly-supervised methods.