Abstract: To address the issue of false detections in dam crack detection caused by low-quality surveillance images, limited effective samples, and interference from complex backgrounds, this study proposes an improved YOLOv8n-based detection method. The model is trained on 193 real-world crack images featuring complex engineering backgrounds. Enhancements include modifications to the mosaic data augmentation mechanism and the incorporation of negative sample training targeting falsely detected objects. Experimental results demonstrate that: (1) under small-sample training conditions, the YOLOv8n model achieves a mean Average Precision (mAP) of 89.2%, meeting the requirements of general engineering applications; (2) after negative sample training, the mAP increases to 92.5%, and the false detection rate is reduced by 10.1%, effectively addressing the false detection problem in complex background scenarios. The findings indicate that the YOLOv8n model is well-suited for dam surveillance images with suboptimal quality, and that the negative sample training strategy significantly improves detection accuracy. This approach offers a novel solution to crack identification in hydraulic engineering and holds substantial value for practical applications.