Abstract: Buoys are affected by wind,waves,weather or impacts,often resulting in tilting,and substantial tilting may cause damage to buoys.The traditional method of buoy inspection mainly relies on manpower,which is not only timeconsuming and labor-intensive,but also the harsh operating conditions threaten the safety of operators.Therefore,designing an intelligent method for detecting damage to navigational markers is of great significance for safeguarding the safety of operators and improving the efficiency of maritime supervision.Considering the limited computational capability of hardware devices in the actual operation on the sea surface,which leads to the problems of slow detection speed and low accuracy,a lightweight buoy tilt detection method based on machine vision is proposed.Firstly,based on the single-stage detection network YOLOv5,CSL（Circular Smoothing Labels） transform the angle regression problem into a classification problem and proposes the R-YOLO buoy tilt detection model.Then.the lightweight GhostNet feature extraction module is added can decrease the model parameters to suit the actual operational environment.Finally,a CBAM（Convolutional Block Attention Module） attention mechanism module can improve the accuracy of the buoy tilt detection model.The experimental results demonstrate the improved tilt detection method in this paper improves the accuracy by 5.6% and reduces the parameter quantity by 2.74×10⁶ in comparison with R-YOLO,and the highest accuracy can reach 93.2%.