To address the issues of overfitting, low computational speed, and insufficient target information extraction in pavement crack segmentation tasks, this study proposed a Swin-U network model based on the U-Net architecture. The model adopted the Swin-Transformer as the feature extraction module to enhance the model’s fitting capability and enable more accurate crack feature extraction, thereby enhancing segmentation accuracy. Additionally, a stable loss function, Focal Loss, was introduced to improve the accuracy of target segmentation. Experiment results on a proprietary pavement crack dataset show that the Swin-U network model achieves pixel-level segmentation of crack images and significantly outperforms the traditional U-Net. On the test set, it improves the intersection over union and F1 score by 25.00% and 27.61%, respectively. This improved model not only provides more reliable technical support for road maintenance decision-making but also offers a reference for the optimization of pavement crack segmentation methods.