基于迁移学习的茶叶嫩芽识别方法

Tea shoot recognition based on transfer learning

摘要

针对当下茶叶生产中普遍存在的人工采摘成本高及传统嫩芽识别技术精度不够高的问题。为了研究出一种具备强泛化能力,适用于不同品种和不同环境的嫩芽识别算法。该研究通过采集大量不同茶树品种、不同光照强度、不同时间段的茶叶图像构建高质量的深度学习训练数据集,选用目标检测神经网络模型CornerNet,YOLOv7和Faster RCNN对一芽一叶进行样本训练和测试,并引入迁移学习(transfer learning,TL)方法优化特征提取与模型收敛,显著提高算法在跨品种、跨环境下的泛化能力。对比分析不同网络模型的识别效果。结果显示,融合迁移学习的TL-YOLOv7网络模型识别一芽一叶的效果最优,不同茶树品种、不同光照条件以及不同时间段下对嫩芽识别AP值达到95.3%,召回率为92.9%,每幅图像的检测时间为0.017 s,实现了高精度和实时性的统一,从而得出TL-YOLOv7网络模型能够有效识别茶叶嫩芽的结论,并在不同条件下具有良好的适应性,为茶叶机械化智能采摘提供了可靠的算法支撑。

Abstract

In response to the current difficulty in tea picking and the insufficient accuracy of traditional sprout recognition.To develop a tender shoot recognition algorithm suitable for different varieties and environments. A deep learning training dataset is constructed by collecting a large number of tea images. Three object detection neural network models, CornerNet, YOLOv7, and Faster RCNN are selected for training and testing one shoot and one leaf samples. Transfer learning methods are incorporated into these models to enhance their generalization ability. The recognition results of different network models are compared and analyzed. The results show that the TL-YOLOv7 network model demonstrated the best performance in identifying one shoot and one leaf. The AP value for identifying tender buds in different varieties, lighting conditions, and time periods reached 95.3%, along with a recall rate of 92.9%. The detection time for each image is 0.017 seconds. Therefore, it can be concluded that the TL-YOLOv7 network model can effectively recognize tea leaf shoot and demonstrate good adaptability under different conditions.

关键词

茶叶嫩芽 / 目标检测 / 深度学习 / 迁移学习 / TL-YOLOv7算法

Key words

tea shoot / object detection / deep learning / transfer learning / TL-YOLOv7 algorithms