基于多尺度与注意力机制的毛尖茶分类及掺假程度  

Classification and adulteration degree of Maojian tea based on multi-scale and attention mechanism

作　　者：毛腾跃[1,2] 伍竞成 MAO Tengyue;WU Jingcheng(South-Central Minzu University College of Computer Science,Wuhan 430074,China;South-Central Minzu University Research Center of Intelligent Management Engineering Technology for Manufacturing Enterprises in Hubei Province,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院,武汉430074 [2]中南民族大学湖北省制造企业智能管理工程技术研究中心,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2024年第6期790-796,共7页Journal of South-Central Minzu University(Natural Science Edition)

基　　金：湖北省科技重大专项(2020AEA011);湖北省高价值知识产权培育工程项目(D2021002094)。

摘　　要：针对消费者在生活中难以区分毛尖茶品种及掺假程度多少的问题,提出了一种基于多尺度特征提取与高效通道注意力机制相结合的网络模型.在DenseNet121的基础上使用多尺度特征提取结构替换原来单一的卷积核,丰富特征层信息,在模型的密集连接块中引入ECA-Net注意力机制,增强有效特征信息的传递,而后,对模型的参数进行调优,进一步提高模型的识别性能.结果表明:改进后的MS-ECA-DenseNet121-C分类模型在收集的8个类别的毛尖种类及掺假种类数据集上的识别准确率达到了96.95%,可以有效鉴别毛尖茶品种的真实性,且改进后的模型大小仅为27.3 MB,便于部署于手机端,在茶叶识别领域具有一定的应用价值.In order to solve the problem that it is difficult for consumers to distinguish the varieties of Maojian tea and the degree of adulteration in their daily life,a network model based on multi-scale feature extraction and efficient channel attention mechanism is proposed.Based on DenseNet121,the multi-scale feature extraction structure is used to replace the original single convolution kernel,which enriches the information of the feature layer.Then,ECA-Net attention mechanism is introduced at the dense connection block of the model to enhance the transmission of effective feature information.Finally,the parameters of the model are optimized to improve its recognition performance.The results show that the improved MS-ECA-DenseNet121-C classification model achieves a recognition accuracy of 96.95%on the collected data sets with eight varieties of Maojian tea and their adulteration,which can effectively identify the authenticity of Maojian tea varieties.And the size of the improved model is only 27.3 MB,which is easy to deploy on the mobile phone and has certain application value in the field of tea recognition.

关 键 词：毛尖茶 密集连接网络 多尺度特征提取 注意力机制 茶叶掺假 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]