Abstract
Fruit detection using the YOLO framework has fostered fruit yield prediction, fruit harvesting automation, fruit quality control, fruit supply chain efficiency, smart fruit farming, labor cost reduction, and consumer convenience. Nevertheless, the factors that affect fruit detectors, such as occlusion, illumination, target dense status, etc., including performance attributes like low accuracy, low speed, and high computation costs, still remain a significant challenge. To solve these problems, a collection of fruit images, termed the CFruit image dataset, was constructed, and the YOLOcF fruit detector was designed. The YOLOcF detector, which is an improved anchor-based YOLOv5, was compared to YOLOv5n, YOLOv7t, YOLOv8n, YOLOv9, YOLOv10n, and YOLOv11n of YOLO variants. The study findings indicate that the computation costs in terms of params and GFLOPs of YOLOcF are lower than those of other YOLO variants, except for YOLOv10n and YOLOv11n. The mAP of YOLOcF is 0.8%, 1.1%, 1.3%, 0.7%, and 0.8% more accurate than YOLOv5n, YOLOv7t, YOLOv8n, YOLOv10n, and YOLOv11n, respectively, but 1.4% less than YOLOv9t. The detection speed of YOLOcF, measured at 323 fps, exceeds that of other YOLO variants. YOLOcF is very robust and reliable compared to other YOLO variants for having the highest R2 of 0.422 value from count analysis. Thus, YOLOcF fruit detector is lightweight for easy mobile device deployment, faster for training, and robust for generalization.