A Multi-ripeness Citrus Detection Model Based on OR-YOLOv11

Efficient detection of fruit is critical for agricultural harvesting robots but remains challenging due to visual occlusion and complex orchard environments. To address the limitations of current detection models, an enhanced multi-ripeness citrus fruits detection model (OR-YOLOv11) is proposed, which is developed based on the You Only Look Once (YOLO). Specifically, the RepConv Ghost CSPELAN (RCGCSPELAN) module enables effective compression of parameter count and model size by integrating the linear transformations of the Ghost module with the feature recombination of the CSP structure. Then, a Deformable Attention (DA) mechanism with adaptive adjustment capability was introduced. This enables the model to focus more effectively on occluded fruit regions, thereby enhancing ability to extract spatial distribution features of citrus fruits. Additionally, a Context-Guided Feature Pyramid Network (CGFPN) was integrated into the architecture. By reorganizing multi-scale features, the module enhances the capability, ultimately improving the model’s precision in complex agricultural scenes. Experiment results demonstrates that OR-YOLOv11 achieves a precision of 87.7%, recall of 79.8%, and mAP@0.5 of 89.8%, with a compact model size of only 5.42 MB, outperforming the baseline YOLOv11n. Furthermore, field tests deployed on a harvesting robot performance with a recall of 75.1% and an inference speed of 19.6 FPS. In conclusion, the proposed model provides a robust and efficient visual perception framework for intelligent citrus harvesting.