基于多尺度注意力机制和知识蒸馏的茶叶嫩芽分级方法

Tea Buds Grading Method Based on Multiscale Attention Mechanism and Knowledge Distillation

相较于人工感官评审法,基于深度学习和计算机技术进行茶叶嫩芽分级可以降低时间成本并大幅提高精度,但常用的识别模型存在着冗余计算量多和模型规格大的问题。为此以采摘自贵州红枫山韵茶场的茶叶嫩芽为研究对象,根据人工经验将茶样划分为3个等级;在ShuffleNet-V20.5x基本单元中嵌入多尺度卷积块注意力模块(MCBAM)与多尺度深度捷径(MDS),提出一种茶叶嫩芽分级模型(ShuffleNet-V20.5x-SMAU),聚焦茶样中有利于分级的特征信息;以在两个不同源域上预训练后的模型作为教师和学生模型,提出一种结合双迁移和知识蒸馏的茶叶嫩芽分级方法,借助暗知识的传授进一步增强分级模型分类性能与抵抗过拟合的能力。结果表明,本文方法能在保证模型轻量性的条件下,对测试集各级样本的分级准确率达到100%、92.70%、89.89%,表现出优于采用复杂网络模型的综合性能,在储存资源有限和硬件水平低的生产场景中应用具有优越性。

Compared with the artificial sensory evaluation method,the tea bud grading based on deep learning and computer technology can reduce the time cost and greatly improve the accuracy,but the commonly used recognition model has the problem of large redundant calculation and large model specifications.For this reason,the tea buds picked from the Hongfeng Mountain Yun Tea Farm in Guizhou were used as the research object,and the tea samples were divided into three grades based on the workers’experience.The multiscale convolutional block attention module(MCBAM)and multiscale depth shortcut(MDS)were embedded in the ShuffleNet-V20.5 x basic unit,a tea bud grading model(ShuffleNet-V20.5 x-SMAU)was proposed,which focused on the feature information in tea samples that was conducive to grading.The models pre-trained on two different source domains was taken as the teacher and student model.A tea bud grading method was proposed which combined dual migration and knowledge distillation.With the help of dark knowledge,the classification performance of the grading model and the ability to resist over-fitting were further enhanced.The results showed that the classification accuracy of the method can achieve 100%,92.70%and 89.89%respectively for the three different grade samples in the test set under the condition of ensuring the lightweight of the model,which was better than the comprehensive performance of the complex network model.The application was more advantageous in production scenarios with limited storage resources and low hardware levels.