CO-YOLO: A lightweight and efficient model for Camellia oleifera fruit object detection and posture determination

The complex growth patterns of Camellia oleifera fruits in natural environments pose significant challenges for harvesting robots. Conventional solutions often rely on complex 3D point cloud processing to detect these growth patterns and enable robotic harvesting. In this study, a method for recognizing the growth patterns of Camellia oleifera fruits in 2D images is proposed. To address the challenges of precise posture detection, the fruits are categorized into five types based on their growth patterns: Front, Up, Down, Left, and Right. Occluded fruits are classified separately, and a dedicated dataset for posture recognition is created. Furthermore, a posture detection model, CO-YOLO, based on the YOLO11n architecture, is introduced. The Multi-scale Aggregation Attention (MMA) module replaces the original C3f2 module, enabling the fusion of feature information across multiple scales, which enhances the model’s perceptual capabilities and improves posture recognition accuracy. Additionally, the Depth Pointwise Convolutional (DPW) module is introduced to replace standard convolutions in the backbone and neck networks, enabling better fusion of channel features, enhancing the representation of posture features in the target region, and reducing the number of parameters. Experimental results show that CO-YOLO achieves a precision of 90.6 %, a recall of 87.0 %, and an [email protected] of 93.7 %. Compared to YOLO11s, CO-YOLO reduces model size and computational complexity by 77.1 % and 69.9 %, respectively, while improving [email protected] by 4.8 %. Compared to YOLOv7-tiny, YOLOv9s, and YOLOv10s, CO-YOLO achieves increases in [email protected] of 3.9 %, 5.6 %, and 5.5 %, respectively. Heatmaps generated by CO-YOLO and YOLO11n indicate that these enhancements significantly improve posture recognition. In summary, the CO-YOLO model exhibits strong performance and provides valuable insights for advancing fruit-picking robotics.