【作者】 赵磊；

【导师】 宋彦；

【作者基本信息】 安徽农业大学 ， 农业工程， 2023， 硕士

【摘要】 茶叶拼配是茶叶精加工中的一项工艺,旨在参考茶叶标准样,选择两种及以上外形、品质不一的原料茶按一定比例拼合,使拼配后茶叶的质量符合产品要求。目的在于提高茶叶品质一致性、稳定茶叶质量、扩大供货来源。在我国茶叶生产过程中,仍依靠拼配专家完成拼配工作,存在主观性强、难以定量等问题,不利于实现茶叶质量标准化。尽管目前理化分析方法可以用于检测茶叶品质,但由于试验过程复杂,检测速度较慢等问题,仍难以在茶厂中实现规模化普及。随着传感器技术的发展,各类图像、波谱检测技术被广泛用于茶叶的品质分析中,这为解决拼配问题提供新的思路。本文以眉茶为研究对象,采用计算机视觉技术与高光谱成像技术,面向拼配工艺中的成品茶等级判别问题、试拼小样与标准样的相似性评价问题、拼配比例设计问题开展研究,取得了以下主要结果:(1)针对眉茶等级判别问题,构建了表征其外形品质的几何特征,结合机器学习方法建立眉茶等级判别模型,实现7个等级眉茶的精准分类。本文提取了10种叶片几何特征并构建描述茶叶外形品质的特征直方图,分别基于XGBoost、LMNN和卷积神经网络3种方法建立外观等级判断模型。结果表明,3种方法在测试集中5折交叉验证的平均准确率均高于92%,其中使用卷积神经网络结合Softmax分类器的方法分类精度最高,测试集平均准确率为97.86%。结果表明,本文构建的外形几何特征能有效表征眉茶外观品质,结合卷积神经网络模型,能够精准实现眉茶等级分类。(2)针对茶叶拼配过程中试拼小样与标准样的相似性评价问题,本文提出了一种基于深度度量学习的相似性评价方法。本文采用7种等级的眉茶标准样作为训练样本,并在标准样中加入不同含量半成品茶构建具有不同相似性的测试样本。采集茶样的高光谱数据并提取光谱特征与图像特征,分别以光谱数据、图像数据、图谱融合数据3种类型的数据作为模型输入。为了构建距离特征空间,本文建立深度度量学习模型并设计了Center Anchor Triplet Loss损失函数,通过样本数据在特征空间中的距离,表征相似程度,达到定性判断相似性和定量度量相似度的目的。结果表明:深度度量学习方法可以准确评价茶样相似性,图谱融合数据结合Center Anchor Triplet Loss的方法精度最高,相似性判断准确率为98.89%,相似度度量准确率为100%。本文采用未经训练的独立样本评价模型,获得了较好的结果,说明算法具有较好的泛化能力,研究结果为眉茶的相似性评价提供了理论依据。(3)为了实现茶叶拼配比例的智能计算,本文构建基于强化学习机制的拼配比例计算模型,求解目标样拼配比例。选取4种眉茶拼配中常用的半成品茶,按比例拼合获得10种品质不同的茶样作为目标样。采集上述茶样的高光谱数据,面向拼配比例计算框架中的品质判断、茶叶光谱生成、拼配方案探索问题,分别构建机器学习模型。针对目标状态判断问题,基于深度度量学习构建茶叶相似性判断网络,用以评价当前状态茶样特征与目标样之间的相似性,判断是否符合目标样品质要求;针对数字化拼配中无法采集每种比例茶叶光谱的问题,基于TF-ACGAN网络训练茶叶光谱生成器,生成光谱被茶叶相似性判断网络准确分类的概率为94.25%;针对眉茶拼配问题设计状态、动作与奖励函数,结合茶叶相似性判断网络与茶叶光谱生成器分别构建GM-Q-learning和GM-TD3拼配比例计算模型,测试模型计算拼配比例的效果。结果表明,基于强化学习的茶叶拼配算法可以准确求解目标茶样的拼配比例,实现茶叶智能化拼配。本文面向拼配工艺中的茶叶等级判别问题,构建外形几何特征描述外观品质,实现茶叶等级分类;面向试拼小样与标准样的相似性评价问题,建立深度度量学习模型,利用样本特征距离量化茶样间的相似度;面向拼配比例计算问题,构建强化学习模型准确计算目标样拼配比例。本文针对拼配过程中的若干决策问题,提出了基于机器学习技术的智能化算法,有助于解决茶叶拼配技术中主观性较强的问题,提高拼配过程中的智能化水平,实现茶叶品质标准化。更多还原

【Abstract】 Tea blending was a process in tea refining aimed at referring to the national standards and standard samples of tea,blending two or more raw teas with different shapes and qualities according to a certain ratio,and making the quality standards of blended tea meet national requirements.The purpose was to improve the consistency of tea quality and stabilize tea quality.In the tea production process in China,blending experts judged whether the trial blend met the national standards through sensory evaluation,which had strong subjectivity and was difficult to quantify,which was not conducive to achieving standardization of tea quality.Although methods such as physical and chemical analysis can be used to detect tea quality,it is still difficult to achieve large-scale popularization in tea factories due to high cost,high time consumption,and inability to detect online.With the development of sensor technology and machine learning technology,computer vision and hyperspectral cameras can accurately characterize the shape and internal characteristics of tea,and machine learning methods can effectively analyze data and mine deep features.This thesis took Mee tea as the research object,based on computer vision technology and hyperspectral imaging technology,and conducted research on the identification of finished tea grades,the evaluation of the similarity between trial blends and standard samples,and the design of intelligent blending schemes in the blending process.The main results of this study were as follows:(1)Aimed at the problem of Mee tea grade discrimination,this study collected and analyzed tea leaf images through computer vision technology,and established a Mee tea grade identification model based on machine learning methods to achieve accurate classification of Mee tea grades.This thesis used digital image processing technology to analyze tea leaf images,extracted 10 external features data such as leaf length,leaf width,leaf area,leaf perimeter,and rectangularity,and constructed a histogram of tea leaf external features.Three methods,XGBoost,LMNN,and convolutional neural network,were used to establish appearance grade judgment models.The results showed that the average accuracy of the three methods in the test set with 5-fold cross-validation was higher than92%,and the method that used convolutional neural network combined with Softmax classifier had the highest classification accuracy,with an average accuracy of 97.86% in the test set.The experiment shows that machine learning methods can learn differentiated information of different categories of tea from the histogram of external features,accurately achieving the classification of eyebrow tea grades.(2)Aimed at the similarity evaluation of the trial sample and the standard sample in the process of tea blending,this thesis proposed a similarity evaluation method based on deep metric learning to evaluate the similarity between the trial blend sample and the standard sample in the tea blending process.Seven levels of standard eyebrow tea samples were used as the training set,and test set with different similarity levels was constructed by adding different proportions of semi-finished tea to the standard samples.The high spectral data of tea samples were collected,and spectral features and image features were obtained.The three types of data,including spectral data,image data,and graph fusion data,were used as inputs to the model.To construct the distance feature space,this thesis proposed a deep feature extraction network based on triplet loss and designed the Center Anchor Triplet Loss function to qualitatively judge the similarity and quantitatively measure the similarity based on the distance of tea spectral data in the feature space.The results showed that the method of graph fusion data combined with Center Anchor Triplet Loss had the highest accuracy,with a similarity judgment accuracy of 98.89% and a similarity metric accuracy of 100%.An independently evaluated model without training was used in this thesis,which achieved good results,indicating that the algorithm had good generalization ability.The research results provide a theoretical basis for the similarity evaluation of Mee tea.(3)In order to achieve the goal of intelligent tea blending,this thesis combined generative adversarial networks,deep metric learning networks,and reinforcement learning networks to construct a tea blending model that provides blending solutions for samples that meet the quality requirements of the target sample.In this experiment,4commonly used semi-finished teas for Mee tea blending were selected,and 10 tea samples with different qualities were obtained by blending them in proportion as target samples.The hyperspectral data of the aforementioned 14 tea samples were collected.Based on deep metric learning,a network was constructed to judge the similarity between the current state tea sample features and the target sample,to determine the distance of the tea target state in order to make a decision on the tea blending.To solve the problem that the spectrum cannot be directly calculated when the tea ratio changes in digital blending,a tea spectrum generator was trained based on the TF-ACGAN network,and the probability that the generated spectrum was accurately classified by the discriminator was 92.50%.In order to design the state,action,and reward functions for the Mee tea blending problem,GM-Q-learning and GM-TD3 networks were constructed based on the tea target state judgment network and the tea spectrum generator to automatically explore tea blending solutions and requested the blending ratio.The results show that the reinforcement learning method can enable the tea matching agent to accurately learn how to achieve the target state from the original state,and realize the intelligent tea matching.This thesis utilizes computer vision and hyperspectral technology to obtain tea characteristics,and builds a model based on machine learning methods to provide a digital approach for tea blending process,which involves tea grade classification,tea similarity evaluation,and blending scheme design.更多还原