A YOLO-based citrus image recognition and harvesting system

To address the challenges of low efficiency and high labor intensity in manual citrus harvesting, this study proposes an intelligent citrus recognition and automated picking system based on YOLOv8. An automated harvesting mechanism was designed, incorporating an image acquisition module, a 6-degree-of-freedom robotic arm, and a picking end-effector assembly. The image acquisition module utilizes a depth camera to capture citrus images and spatial coordinates, while the robotic arm positions the end-effector to execute the harvesting operation. The YOLOv8 model was employed for robust citrus detection. Experimental results demonstrate a detection confidence threshold ≥0.82, with high precision maintained across most confidence intervals – achieving perfect precision (1.0) at a confidence level of 0.914. While recall gradually decreased with increasing confidence thresholds, it remained at relatively high levels overall, yielding a mean average precision (mAP@50) of 0.852. Optimal balance between precision and recall was observed at confidence thresholds of 0.1-0.7, with corresponding F1-scores ranging from 0.7 to 0.82. Ablation studies on illumination invariance confirmed the necessity of color enhancement, with both high-intensity and low-intensity light enhancement groups showing significant performance improvements under specific conditions, albeit with potential overfitting risks to particular lighting domains. This research provides an effective technical solution for vision-based citrus recognition and automated harvesting in natural orchard environments.