RTD-YOLO: a precise quality detection and grading model for rose tea

Against the backdrop of the increasingly prominent commercial value of roses, rose tea with uneven quality levels has no competitive advantage. To enhance the market competitiveness of rose tea, it is necessary to classify it. At present, the grading of rose tea mainly relies on manual labor, which is inefficient. In addition, there may be difficulties in deploying large models in actual production. Therefore, we proposed a model for quality detection and grading of rose tea based on YOLOv12n. Firstly, the SEFF module was integrated into the A2C2f module to replace the A2C2f module in neck. Then, ADown was used to replace Conv. Thirdly, a parameter sharing detection head was constructed. Finally, the model was pruned using the Lamp method. The FLOPs, Params, F1, mAP, FPS, and Model size of the improved model are 2.6 G, 0.83 M, 90.48%, 95.85%, 298.5, and 2.0MB, respectively. Compared with the original model, F1, mAP, and FPS have increased by 2.93%, 1.55%, and 20.5, respectively, while FLOPs, Params, and Model size have decreased by 3.2 G, 1.68 M, and 3.2MB, respectively. The model supports real-time deployment at 55.3 FPS on Jetson Orin NX edge devices. The improved model provides technical support for intelligent grading of rose tea.