Research on target recognition method for machine-plucked fresh tea leaves based on improved YOLOv11.

In the process of multi-stage singulation, machine-plucked fresh tea leaves are arranged in a low-occlusion manner; however, challenges such as partial occlusion, difficulty in small-object detection, and the need for accurate recognition and classification remain in sorting. To address these issues, this study proposes an improved YOLOv11-based detection model, termed YOLOv11-MSC. The model incorporates three key enhancements-multi-scale edge information enhancement, convolution-attention fusion, and spatial attention optimization-to strengthen feature extraction and localization capabilities. To comprehensively evaluate its effectiveness, a series of experiments were conducted, including ablation studies, comparative analyses with different attention mechanisms, comparisons with mainstream detection models, and visual examinations using feature maps and heatmaps. Experimental results demonstrate that YOLOv11-MSC achieves a mean average precision of 90.1% (mAP@0.5) while maintaining a lightweight architecture with only 2.89 M parameters and 6.4 GFLOPs. These results demonstrate excellent performance in both detection accuracy and computational efficiency, indicating strong suitability for deployment in real-time tea-leaf grading applications.