Efficient tea bud recognition is vitally important to improve the harvesting performance of premium tea. Numerous studies have been directed toward the development of recognition models. To overcome the computational limitations of edge devices without compromising recognition precision in complex tea garden environments, a recognition model entitled YOLO-ECN is proposed in this study. This model adopts the EfficientnetV2-S network as its backbone, employing a hierarchical lightweight design and a compound scaling strategy to simultaneously optimize the accuracy and parameter efficiency. To mitigate the missed detection, the Convolution and Attention Fusion Module (CAFM) is developed by incorporating an adaptive dynamic weighting strategy to strengthen the feature representation. Moreover, a hybrid loss function integrating the Normalized Wasserstein Distance and intersection over union is developed to improve the capture ability of occluded tea buds. Ablation experiments verify that the proposed YOLO-ECN achieves a precision of 88.4%, recall of 88.6% and mAP of 88.9%, which surpasses the original YOLOv10s model by 9.5 percentage points in precision, 4.4 percentage points in recall and 10.4 percentage points in mAP. The YOLO-ECN demonstrates a competitive combination of high accuracy and low computational cost compared to the mainstream models of Faster R-CNN, Tea-YOLO, YOLOv7MCS and YOLOv11n, providing an efficient recognition model for the automated picking of tea buds.
Pan et al. (Sun,) studied this question.