Aiming at the problem that hydraulic structures are prone to deterioration such as water stains and calcium precipitation under complex natural environments and operating conditions, thereby complicating surface crack detection, this paper proposes a crack detection model YOLO-SEH integrating a deformable convolutional network (DCNv2) and scaling efficient decoupled head (SEH). An efficient channel attention mechanism is introduced into the model neck to suppress noise interference from complex backgrounds. In the backbone, DCNv2 is adopted to replace the C3 module of YOLOv5, improving the model’s adaptability to irregular crack shapes. SEH is incorporated into the detection head to strengthen the model’s sensitivity and discriminability to crack features. The proposed model was applied to crack detection on sluice and dam images captured by unmanned aerial vehicles (UAVs), achieving mAP@0.5 values of 86.2% and 84.9%, respectively. Ablation experiments show that compared with the YOLOv5 model, the mAP@0.5 of the YOLO-SEH model increased by 3.1 and 4.1 percentage points, the precision increased by 4.8 and 6.6 percentage points, and the recall increased by 3.4 and 4.7 percentage points for the two datasets, respectively. Furthermore, comparative experiments with the Faster R-CNN, SSD300, YOLOv5, YOLOv7, and YOLOv8 models demonstrate that the YOLO-SEH model achieves superior detection performance for crack detection of hydraulic structures.