Crack segmentation is a critical task in the maintenance of modern civil infrastructure, where rapid and accurate crack detection is essential. In view of the limitations of convolution-based models in long-range modeling and the flaws of Transformer-based models in local feature extraction and high computational complexity, a frequency-enhanced U-shaped network (FE-UNet) has thus been proposed. First, a frequency-enhanced attention module (FEA) is introduced, which adopts fast Fourier transform to map the image from the spatial domain to the frequency domain, thus reducing computational complexity. A convolution module is subsequently introduced in the frequency domain to highlight crack specific frequency components, thereby enhancing feature representation. In addition, a local-enhanced feed-forward network (LE-FFN) is proposed to enhance the local interaction capability of features by integrating multi-scale semantic information from multiple branches. Experimental results show that the proposed network achieves a maximum mIoU of 82.38% and F1 of 79.48%, providing strong support for road health monitoring.