An end-to-end detection and classification model for tea leaf grading in complex orchard environments

Existing detection models for automatic tea-grade determination in open-field, complex habitats suffer from insufficient feature robustness, weak suppression of background interference, and difficulty in balancing lightweight design with accuracy. To address these limitations, this study proposes Tea-DETR, a model tailored for open-air tea garden scenarios based on the end-to-end RT-DETR detection paradigm.

The proposed method enhances feature representation and attention allocation from two aspects: backbone optimization and feature interaction enhancement. First, a lightweight backbone module is designed to strengthen long-range dependency modeling and improve suppression of complex background interference, thereby enhancing the robustness and discriminability of multi-scale tea-leaf features under occlusion and illumination variations. Second, an efficient attention-enhancement mechanism is introduced to reduce redundant feature interactions and enable adaptive focus on critical target regions, improving the model's ability to capture subtle semantic differences among tea grades while maintaining computational efficiency. To validate the approach, Anji white tea from Zhejiang Province was selected as the experimental subject, and a single-leaf tea dataset containing 6,542 images was constructed. The dataset was split into training, validation, and test sets at a ratio of 8:1:1, based on which comparative experiments and convergence analyses were conducted.

With the number of parameters reduced to 14.65M, Tea-DETR achieves an accuracy of 92.2% and improves mAP@0.5 to 82.0%, reducing parameters by 26.8% compared with the baseline model. In addition, Tea-DETR exhibits markedly improved convergence stability and substantially enhanced capability in distinguishing subtle semantic differences among tea grades, effectively alleviating ambiguous discrimination under complex backgrounds.

Overall, the proposed method enhances the stability of fine-grained feature capture for small tea-leaf targets in open-field environments while maintaining real-time inference efficiency, providing a robust solution for real-time, non-destructive automatic tea-grade determination in field scenarios.