YOLO11-ECA: A Lightweight Tea Bud Detection Method for Complex Tea Plantation Environments

Accurate tea bud detection in natural plantation environments is challenged by variable illumination, foliage occlusion, target-background similarity, and dense target distribution. This study proposes YOLO11-ECA, a lightweight detector derived from YOLO11n, to improve the detection performance while reducing the computational demands. The P3/8 prediction branch and its associated high-resolution feature fusion path are removed, while the P4/16 and P5/32 detection scales are retained. Standard adaptive efficient channel attention (ECA) modules are then integrated into the retained feature paths for lightweight channel recalibration. On the validation set, YOLO11-ECA achieved 79.8% precision, 85.1% recall, an 89.2% mAP@0.5, a 52.8% mAP@0.5:0.95, and an F1 score of 82.0%. Relative to YOLO11n, these metrics increased by 1.9, 2.6, 3.1, 0.4, and 2.0 percentage points, respectively, while the parameters and GFLOPs decreased by 7.4% and 28.8%. Scale-specific evaluation showed that P3 removal reduced the AP_S from 36.0% to 32.4%, whereas adding ECA to the dual-scale architecture increased the AP_S to 37.0%. On an independent test set, YOLO11-ECA achieved 82.2% precision, an 84.8% mAP@0.5, and an F1 score of 81.0%. The model also reached 86.28 FPS with end-to-end latency of 11.59±0.75 ms under a unified GPU evaluation protocol. These results indicate a favorable balance between tea bud detection performance and computational efficiency.