Detecting road defects is crucial for maintaining road safety and prolonging pavement lifespan. Significant progress has been made in the visual detection of road defects using convolutional neural networks (CNNs). However, existing CNN-based methods perform well in capturing local features, but the stacked convolutional layers have some limitations in modeling long-range dependencies. In contrast, Transformers excel in capturing global context and long-range dependencies, and they fall short in handling local details, which limits their effectiveness in complex scenarios. To overcome this challenge, this study proposes a Transformer-embedded multiscale framework based UNet (TEM-UNet) that adopts a strategy integrating multi-scale features from both CNN and Transformer models. Using CNN to extract local features of road surface defects, leveraging Transformers to capture global features of defect images, and integrating features at multiple scales to model defect segmentation results simultaneously. The method of integrating multi-scale features not only preserves the advantages of CNN in handling details but also harnesses the powerful capability of Transformers in capturing long-range dependencies. By validating our method on one real-world crack dataset, it demonstrated superior segmentation performance and better generalization ability. Compared to traditional CNN methods and other advanced pavement distress detection algorithms, our approach exhibited higher detection accuracy and robustness.