<p>Cracks pose a serious threat to tunnel structures’ durability and safety, so accurate crack recognition and quantitative measurement are critical for ensuring tunnel operation and maintenance safety. In order to solve the issues of insufficient segmentation accuracy brought on by the variable characteristics of cracks and severe background interference in tunnel environments, this study proposes Crack-YOLOv11, an instance segmentation model based on an improved YOLOv11. The original backbone network is initially replaced with the EfficientNetV2-S structure, which integrates the SimAM attention mechanism, in order to enhance key expression capabilities and extract rich contextual information. Second, to capture features more typical of cracks, the 2D dynamic snake convolution module is incorporated at the neck of the model. Additionally, addressing the coupling between detection and segmentation in instance segmentation, an adaptive task weighting scheme optimizes the loss function for more effective crack segmentation training. Finally, pixel scale factors calculated from tunnel ring seams are combined with crack segmentation masks to perform quantitative measurements on crack recognition. According to experiments, Crack-YOLOv11 performs 11.7%, 4.6%, and 4.8% better in mAP@50, F1-score, and IoU than the baseline YOLOv11 model. Robustness testing and measurement results further show that our model maintains high stability and engineering practicality in challenging scenarios, meeting the practical needs of intelligent tunnel inspection.</p>