To address the low segmentation accuracy caused by significant dimensional variations in road and bridge cracks, an enhanced multi-scale crack segmentation network named MSC-YOLO, based on YOLOv8, is proposed. The multi-scale cross-axis attention mechanism captures cross-scale critical features to enhance crack feature perception; dynamic upsampling improves segmentation flexibility for varying crack sizes; and the EIOU loss function optimizes crack boundary fitting. Experimental results on the Roboflow Crack-Seg dataset demonstrate a 6.7% increase in mAP50 compared to the baseline YOLOv8, while maintaining a low parameter volume. This approach enables high-precision severity assessment in resource-constrained environments and is deployable in intelligent infrastructure inspection systems.