An anchor-based YOLO fruit detector developed on YOLOv5

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• YOLO framework enhances fruit yield prediction, automation, and supply chain efficiency.
• Challenges in fruit detection include occlusion, illumination, and computational demands, impacting accuracy and speed.

Purpose of the Study:

• To develop an improved fruit detection model addressing existing challenges.
• To evaluate the performance of the proposed YOLOcF detector against various YOLO variants.

Main Methods:

• Construction of the CFruit image dataset.
• Design and implementation of the YOLOcF fruit detector, an enhanced YOLOv5 variant.
• Comparative analysis with YOLOv5n, YOLOv7t, YOLOv8n, YOLOv9, YOLOv10n, and YOLOv11n.

Main Results:

• YOLOcF demonstrates lower computation costs (params, GFLOPs) than most YOLO variants, except YOLOv10n and YOLOv11n.
• Achieved higher mean Average Precision (mAP) than YOLOv5n, YOLOv7t, YOLOv8n, YOLOv10n, and YOLOv11n.
• Exhibits superior detection speed at 323 fps and highest R2 value (0.422) for robustness and reliability.

Conclusions:

• YOLOcF is a lightweight and robust fruit detector suitable for mobile deployment.
• The model offers faster training and better generalization capabilities.
• Addresses key challenges in fruit detection for smart farming and automated agricultural processes.