Abstract: In order to realize the rapid recognition of tea buds and the location of picking points, a lightweight deep learning network is studied to realize the segmentation of tea buds and the location of picking points. The combination of MobileNetV2 backbone network and dilated convolution can better balance the contradiction between the speed and accuracy of tea bud image segmentation, and meet the requirements of fast recognition of tea buds while achieving high segmentation accuracy. A picking point location method combining outer contour scanning and area threshold filtering is designed. The experiments show that the tea bud segmentation algorithm proposed in this paper has excellent accuracy in single bud tip and one bud one leafdataset, and mIoU reaches 91.65% and 91.36% respectively. While maintaining high accuracy, the model complexity of this algorithm is the lowest, with only 5.81 M parameters and 39.78 GFLOPs calculations. In the single bud tip, one bud and oneleaf, and onebud and twoleave data sets,  200 pictures were randomly selected to verify  the location of picking point, and the positioning accuracy reached 90.38%, 95.26% and 96.60% respectively.

Key words: tea bud, deep learning, semantic segmentation, dilated convolution, receptive field, picking point positioning

摘要： 为实现茶嫩芽快速识别与采摘点定位，研究一种轻量级深度学习网络实现茶嫩芽分割与采摘点定位。采用MobileNetV2主干网络与空洞卷积相结合，较好地平衡茶嫩芽图像分割速度与精度的矛盾，实现较高分割精度的同时，满足茶嫩芽快速识别的要求，并设计外轮廓扫描与面积阈值过滤相结合的采摘点定位方法。试验表明：所提出的茶嫩芽分割算法在单芽尖及一芽一叶数据集中精度优异，平均交并比mIoU分别达到91.65%和91.36%；在保持高精度的同时，模型复杂度低，参数量仅5.81 M、计算量仅39.78 GFOLPs；在单芽尖、一芽一叶及一芽两叶数据集中各随机抽取200张图片进行采摘点定位验证，定位准确率分别达到90.38%、95.26%和96.60%。

关键词: 茶嫩芽, 深度学习, 语义分割, 空洞卷积, 感受野, 采摘点定位