Tunnel structures are prone to surface defects such as water seepage, cracks, and spalling during long-term service, seriously threatening driving safety and structural durability. Existing tunnel defect detection methods still lack multi-scale feature extraction and adaptability to embedded deployment. To address these issues, an improved tunnel-defect detection method named YOLOv11-WTseg was proposed. Specifically, a wavelet transform convolution block (WTConv) was introduced into the segmentation branch of the YOLOv11-seg architecture to enlarge the receptive field and enhance multi-scale feature representation. Meanwhile, the bounding box regression loss was replaced from complete intersection over union (CIoU) to Scylla intersection over union (SIoU) to optimize the regression process through an angle-based penalty mechanism. Based on a self-built tunnel defect dataset, comparative experiments were conducted on gutter damage, non-gutter slab damage, pavement cracks, water seepage, and vault spalling. Results show that the introduction of WTConv increases Box_mAP from 0.821 to 0.844, and Mask_mAP from 0.787 to 0.820. Further employing SIoU increases Box_mAP to 0.845 and Mask_mAP to 0.839. Compared to the baseline model, the improved model demonstrates superior detection and segmentation accuracy across various input scales. Furthermore, embedded deployment verification on the NVIDIA Jetson Xavier NX platform demonstrates that the improved model achieves both lightweight and real-time performance while maintaining a model size of only 5.4 M, achieving an inference latency of 117 ms. These results demonstrate that the YOLOv11-WTseg model can efficiently and accurately detect and segment typical tunnel defects and has promising engineering application prospects, providing technical support for tunnel structural health monitoring and maintenance decision-making.