Grape bunch detection in low-light conditions based on multimodal

Grape bunch detection in low-light night conditions may encounter unexpected situations in the actual production process. Traditional grape bunch picking relies on manual hand-picking, which is inefficient, and in general, visible light is only utilized for the detection of grape bunches in low-light night conditions, but the results are unsatisfactory. In this paper we used both visible light and near infrared modal data, combined the advantages of the two modalities, adopted the feature level fusion of YOLOv11, introduced the Poolformer in Metaformer to improve the C3K2 module,mAP@50 reached 93.3%, Recall reached 87.3%, compared to YOLOv5n Recall improved by 17.3%, mAP@50 improved by 21.1%, compared with YOLOv8n Recall improved by 18.8%, mAP@50 improved by 15.4%, compared with YOLOv10n Recall improved by 26.8%, mAP@50 improved by 20.1%, YOLOv11 Recall improved by 20.0%, mAP@50 improved by 16.6%, and FPS can reaches 340,which can detect grape bunches more accurately in real time.