MRCD-KC: Multi-Role Collaborative Distillation Method for Kiwifruit Counting in Complex Environments

Accurate kiwifruit counting is a core technical prerequisite for smart orchards to achieve automated fruit thinning, intelligent harvesting, and yield prediction. However, factors such as small fruit size, severe occlusion, and scale variation in complex natural environments limit further improvement in counting accuracy. This study proposes a kiwifruit counting method based on multi-role collaborative distillation, which constructs a heterogeneous knowledge-transfer network architecture comprising one teacher, two distinct teaching assistants, and one student, thereby achieving model light-weighting while maintaining detection accuracy. Built upon the YOLOv11 architecture, the teacher network incorporates an Adaptive Occlusion-aware Module (AOM) to enhance the recognition of occluded targets; Teaching Assistant I employs a Multi-Scale Fusion (MSF) module to optimize scale-related features, while Teaching Assistant II integrates a Small-target Enhancement Feature (SEF) module to improve the detection of dense small objects; the student network, based on the standard YOLOv11, fuses complementary knowledge through multi-source feature distillation and self-distillation mechanisms. Experimental results show that the proposed method achieves a precision of 90.30%, a recall of 84.30%, and an [email protected] of 90.70%, all surpassing those of the YOLOv11n baseline, thereby validating its effectiveness in scenarios involving occlusion, scale variation, and dense small object detection. This study not only enriches the theoretical framework for multi-role collaborative distillation in agricultural vision but also provides practical, effective technical support for kiwifruit counting in complex environments.