Abstract: In order to improve the accuracy of tea bud recognition, improve the efficiency of automatic picking robot and reduce the cost of manual picking, this paper proposes a model for tea bud target detection. Through taking pictures of the buds of Baihao early tea, 179 images were obtained after screening, and 716 images were obtained after using Mosc data amplification. The data set was established, and the data set was divided according to the 7∶2∶1 ratio of training set, test set and validation set. In view of the low recognition accuracy caused by the overlap and occlusion of tea buds under complex background, this paper modifies the YOLOv5s model and adds the attention mechanism module SE and CBAM to the backbone network for comparison. The Neck network is changed from the original PAFPN to the BiFPN that can carry out twoway weighted fusion. The Head structure adds a P2 module for shallow sampling, and proposes a tea bud detection model. The experiment shows that the model has higher detection accuracy when YOLOv5s adds SE module combined with BiFPN, and the experimental results are crossverified with ten folds. Compared with the baseline accuracy, the accuracy rate is increased by 10.46%, reaching 88.30%, and the average accuracy mAP is increased by 6.47%, reaching 85.83%. Finally, using the same data set and preprocessing method to compare YOLOv5m, Faster RCNN and YOLOv4tiny, it is proved that the experimental method proposed in this paper is more comprehensive than other classical deep learning methods, can more effectively improve the accuracy of tea bud detection and can provide theoretical basis for the tea automatic picker.

Key words: tea, bud detection, YOLOv5s, attention mechanism, bidirectional feature pyramid

摘要： 为提高对茶叶嫩芽识别的准确率，提升自动采摘机器人的工作效率，减少人工采摘成本，提出一种对茶叶嫩芽目标检测的模型。通过拍摄包含白豪早茶叶嫩芽图片，进行筛选后得到179张图像，使用Mosic数据扩增后获得716张图像，建立数据集，按照训练集、测试集和验证集7∶2∶1的比例划分数据集。针对复杂背景下茶叶嫩芽存在重叠以及遮挡所导致的识别精准度低的问题，对YOLOv5s模型进行改动，在骨干网络上增添注意力机制模块SE和CBAM进行比较；Neck网络由原来的PAFPN改为可以进行双向加权融合的BiFPN，Head结构增加浅层下采样的P2模块，提出一种茶叶嫩芽检测的模型。试验表明YOLOv5s添加SE模块结合BiFPN时模型具有更高的检测精度，并对试验结果进行十折交叉验证，相较于基线精确率提高10.46%，达到88.30%，平均精度均值mAP提高6.47%, 达到85.83%。最后使用相同的数据集和预处理方法对比YOLOv5m、Faster RCNN 和 YOLOv4tiny，证明该试验方法综合强于其他经典深度学习方法，能更有效地提升茶叶嫩芽检测精准度，可以为茶叶自动采摘机器人提供理论依据。

关键词: 茶叶, 嫩芽检测, YOLOv5s, 注意力机制, 双向特征金字塔