Method for lightweight litchi fruit detection in complex environments via improved YOLO11n-MAS

Accurate and real-time detection of litchi fruit is crucial for the development of automated harvesting systems and the advancement of precision agriculture. This study addresses the challenge of litchi fruit detection in complex orchard environments characterized by branch occlusion, fruit overlap, and variable lighting conditions. A lightweight detection model, YOLO11n-MAS, was proposed to achieve efficient and robust fruit identification. A self-constructed Zengcheng litchi dataset was first established to provide high-quality data for training and validation, thereby improving the model’s detection performance. The YOLO11n-MAS model was built on the YOLO11n architecture, incorporating a novel C3k2-EMBC module and a Bi-Level Routing Attention (BRA) mechanism to reconstruct the backbone network, utilizing a parameter-efficient Slimneck to optimize the neck, and introducing a specialized Detect_LSDECD module into the detection head to balance performance and efficiency. Experimental results show that the YOLO11n-MAS model achieves a mean average precision (mAP50) of 0.916, outperforming other advanced lightweight models including YOLOv8n and the baseline YOLO11n. Additionally, the model’s parameter size is only 2.34 M, and its GFLOPs is 5.5, representing reductions of 9.4% and 12.7%, respectively, compared to YOLO11n, significantly lowering computational costs while maintaining high detection performance. The results demonstrate the effectiveness of the C3k2-EMBC, BRA, Slimneck, and Detect_LSDECD modules. The development of the YOLO11n-MAS model can provide an efficient and lightweight solution for litchi fruit detection in resource-constrained environments, offering strong technical support for intelligent harvesting systems.