Grape leafhoppers (Erythroneura spp.) are major pests in grape cultivation, leading to significant economic losses. Accurate and efficient damage level assessment is crucial for effective pest management and reducing financial impact. In this study, we categorized damage into uninfested leaves and five damage levels (I–V) and constructed a grape leafhopper damage dataset. Based on this dataset, we developed a lightweight detection model for grape leafhopper damage levels, incorporating improvements to the YOLOv8n architecture. The model employs FasterNet as the backbone network to enhance computational efficiency and reduce model complexity. It substitutes for the nearest-neighbor upsampling with CARAFE to improve small target detection capabilities. Additionally, the SE attention mechanism is integrated to optimize leaf feature extraction, thereby enhancing recognition accuracy in complex vineyard environments. The experimental results demonstrate that CSF-YOLO achieves a mAP of 90.15%, which is 1.82% higher than the baseline model, YOLOv8n. Additionally, the model’s inference results can be accessed via mobile devices, demonstrating the feasibility of real-time vineyard pest monitoring. This study provides a solid technical foundation for advancing intelligent pest monitoring systems in vineyards and the development of smart agriculture.