Computer vision-based structural monitoring technology is widely applied in intelligent crack identification. The sample scarcity problem is commonly faced in engineering sites. Model performance is limited by traditional machine learning methods under small-sample conditions. A lightweight and real-time crack detection system was constructed based on the YOLOv8 architecture. A hybrid data augmentation method and a k-fold cross-validation strategy were integrated into the system. An initial small-sample dataset of 400 images was built. The dataset was expanded to 3,600 images via data augmentation. A lightweight crack segmentation model was trained. Meanwhile, a visual auxiliary detection interface was developed based on PyQt5. The intuitive display and coordinate output of detection results are realized. Engineering deployment applications and further post-processing operations are facilitated. The Precision, Recall, and mAP@0.5 are evaluated as 0.879, 0.918, and 0.891, respectively. These metrics are improved by 17.7%, 23.7%, and 17.5% compared to Faster R-CNN. These metrics are also increased by 7.7%, 13.2%, and 7.5% compared to YOLOv5. More accurate boundary localization and lower missed detection rates are demonstrated in complex scenarios. A feasible technical solution is provided for intelligent structural monitoring under small-sample constraints. Significant engineering application potential is shown in this research.