<p>Aiming at the problems of poor continuity and low recognition rate of tunnel lining crack segmentation under the influence of seam linear interference and multi-scale differences, and the large number of model parameters, which makes it difficult to balance the efficiency and accuracy of the model well, a tunnel lining crack segmentation method based on the improved UNet architecture, EFE-UNet (Efficient Feature Enhancement UNet), is proposed. Initially, an EMO (Efficient MOdel) lightweight network is used to replace the VGGNet backbone architecture in the original UNet architecture, which enhances the ability to describe the feature context information of the cracks and reduces the parameters of the network at the same time; BRA (Bi-Level Routing Attention) is added at the end of the encoder to implement the low-level feature selection and filter the crack information using the weights, to enhance the model’s accuracy of crack segmentation. Seam linear information to enhance the sensitivity of the model to the crack region and the anti-interference ability to the seam. A DSC (Depthwise Separable Convolution) and Ghost module cascaded with a DDG (Double DSC Ghost) module are used instead of the standard convolution of the decoder to reduce the model complexity and enhance the adaptability to cracks at different scales. On the Tunnel200 dataset, EFE-UNet outperforms other semantic segmentation methods with the best metrics. It also shows great performance on the CRACK500 dataset.</p>