CV-YOLOv10-AR-M: Foreign Object Detection in Pu-Erh Tea Based on Five-Fold Cross-Validation

To address the problem of detecting foreign bodies in Pu-erh tea, this study proposes an intelligent detection method based on an improved YOLOv10 network. By introducing the MPDIoU loss function, the YOLOv10 network is optimized to effectively enhance the positioning accuracy of the model in complex background and improve detection of small target foreign objects. Using AssemFormer to optimize the structure, the network's ability to perceive small target foreign objects and its ability to process global information are improved. By introducing the Rectangular Self-Calibrated Module, the prediction accuracy of the bounding box is effectively optimized, further improving the classification and target-positioning abilities of the model in complex scenes. The results showed that the Box, Cls, and Dfl loss functions of the CV-YOLOv10-AR-M network in the One-to-Many Head task were, respectively, 14.60%, 19.74%, and 20.15% lower than those of the YOLOv10 network. In the One-to-One Head task, they decreased by 10.42%, 29.11%, and 20.15%, respectively. Compared with the original YOLOv10 network, the accuracy, recall rate, and mAP of the CV-YOLOv10-AR-M network were increased by 5.35%, 11.72% and 8.32%, respectively. The CV-YOLOv10-AR-M network effectively improves the model's attention to small sizes, complex backgrounds, and detailed information, providing effective technical support for intelligent quality control in the agricultural field.