Optimized YOLOv5s-based Detection of Cherry Ripeness

Accurate assessment of cherry maturity is crucial for automated harvesting. Current detection technologies are costly, and multi-stage detection models, though detailed, consume significant computational resources and operate slowly, making them unsuitable for weak network or embedded environments typical of intelligent harvesting systems. Single-stage detection models often lack precision in complex environments. To address these issues, this paper introduces the YOLOv5-BE model, which enhances feature fusion capability through the integration of a Bidirectional Feature Pyramid Network (BiFPN) and replaces traditional IOU Loss with EIOU Loss, significantly improving detection accuracy for small targets. Performance evaluations demonstrate that compared to YOLOv5s, YOLOv5-BE increases the mAP@0.5 metric by $5.1 \%$. Compared to other multi-stage models, YOLOv5-BE exhibits superior performance in terms of average precision, operational efficiency, and computational load, proving its effectiveness in intelligent cherry harvesting systems.