In response to the challenges posed by the large number of small targets, complex backgrounds and significant computational load involved in detecting olives, this study presents YOLO-TinyFuse, a lightweight detection model developed based on YOLOv8n. This model incorporates the P2 high-resolution feature layer, a ModifiedNeck cross-scale fusion structure (ModifiedNeck) and a bidirectional feature pyramid network (BiFPN) dynamic weighting module within a unified architecture. This architecture simultaneously preserves high-resolution feature representations, enhances bidirectional multi-scale interaction and optimises weighted feature aggregation. This synergistic design substantially improves the recognition of small objects while reducing model complexity further. Evaluations conducted on a multi-scenario olive phenotyping dataset demonstrate that YOLO-TinyFuse achieves an mAP50 of 92.3% and a Recall of 84.5%. This represents improvements of 2.6% and 3.2% respectively over YOLOv8n, while reducing the parameter count by 6.76%. These results confirm that the proposed model provides a deployable, computationally efficient, real-time solution for target recognition on mainstream edge computing platforms in automated olive harvesting scenarios, and offers a reusable, lightweight framework for agricultural small-object detection tasks requiring high performance and optimised computational efficiency.
Yang et al. (Tue,) studied this question.