Comparative Study of Different YOLO Variants for Ripeness Detection of Palm Oil Fresh Fruit Bunches

Palm oil Fresh Fruit Bunches (FFBs) ripeness detection is the key to precision agriculture, allowing ideal harvesting as well as post-harvesting handling. Multiple deep-learning methods for palm oil FFB ripeness detection were reviewed in the literature. However, the current research challenge is training the deep-learning models using images collected from plantations directly, and hence, detecting palm oil FFBs in real-time with reasonably sound detection accuracies. Hence, this study explored three You Only Look Once (YOLO) algorithm versions to address this challenge: YOLOv5, YOLOv6, and YOLOv7. Notably, no prior research has utilized at least three YOLO versions to recognize the ripeness levels of palm oil FFBs in real-time. This research aims to identify the most effective algorithm among these YOLO versions for this specific task. The methods involved creating a model utilizing a custom dataset as well as conducting training and testing. A custom dataset of 2000 images was obtained from a palm oil plantation in the Selangor region of Malaysia to train and test models. The dataset was trained with YOLOv5s (small), YOLOv5m (medium), YOLOv6s (small), YOLOv6m (medium), and YOLOv7 models. The YOLOv5s model proved the most effective for real-time detection of palm oil FFBs, with a 98.5% F1-score and a solid 35.76 Frames per Second (FPS). Finally, performance analysis showed that the YOLOv5s model outperformed existing methods, offering a dependable solution for palm oil FFB ripening detection in real-time scenarios right at the locations of palm oil plantations.