ECS-tea: a bio-inspired high-precision detection and localization algorithm for young shoots of Pu-erh tea.

Introduction: Pu-erh tea, valued for its ecological significance and economic worth, requires precise and efficient bud harvesting to advance intelligent agricultural operations. Accurate bud recognition and localization in complex natural environments remain critical challenges for automated harvesting systems. Methods: To address this, we propose ECS-Tea, a bio-inspired and lightweight detection-localization framework based on YOLOv11-Pose, tailored for Pu-erh tea bud analysis. The framework integrates four key modules: (1) a lightweight EfficientNetV2 backbone for efficient feature representation; (2) a Cross-Scale Feature Fusion (CSFF) module to strengthen multi-scale contextual information; (3) a Spatial-Channel Synergistic Attention (SCSA) mechanism for fine-grained keypoint feature modeling; and (4) an adaptive multi-frame depth fusion strategy to enhance 3D localization precision and robustness. ECS-Tea was trained and validated on a dedicated dataset for Pu-erh tea bud detection. Results: Experimental results show that ECS-Tea achieves 98.7% target detection accuracy and 95.3% keypoint detection accuracy, with a compact architecture (3.3 MB), low computational cost (4.5 GFLOPs), and high inference speed (370.4 FPS). Compared to the baseline YOLOv11-Pose, ECS-Tea significantly improves keypoint detection performance: mAP@0.5(K) increases by 4.9%, recall R(K) by 3.8%, and precision P(K) by 3.4%, while maintaining or slightly enhancing object detection metrics. Discussion: These findings demonstrate that ECS-Tea effectively balances accuracy and computational efficiency, validating the complementary contributions of its integrated modules. As a robust, real-time, and deployable solution, it bridges the gap between algorithmic sophistication and practical application, enabling high-precision tea bud harvesting in unstructured field environments.