基于近红外光谱和BYOL对比学习的烟叶部位识别方法

Tobacco leaf part recognition method based on near infrared spectroscopy and BYOL contrastive learning

【背景和目的】烟叶部位识别对卷烟制品的配方设计与质量监控具有重要意义。利用近红外光谱(NIR)分析可以实现烟叶部位的快速、无损在线识别。针对烟叶光谱特征提取困难问题,利用具有强特征提取的BYOL模型,提出NIR-BYOL烟叶部位识别方法。【方法】通过微分光谱融合实现数据增强,利用卷积自编码器和多层感知器实现BYOL的在线网络和目标网络,以在线网络和目标网络输出的均方误差为损失函数,通过损失最小优化的编码值,提取的特征经SVM分类识别烟叶部位信息。实验比较分析了不同数据增强方式、卷积核大小和激活函数对模型的影响。【结果】一阶微分融合和二阶微分融合的组合是最佳数据增强方法,对比学习模型最佳参数为卷积核11*1,激活函数为ELU。模型对部位的平均识别率达到91.79%。相比SVM、PCA+SVM和PLS-DA方法,NIR-BYOL模型的准确率有显著提升,分别提升了13.12%、15.79%、16.79%。【结论】近红外光谱分析技术结合对比学习模型可以有效分类识别烟叶的部位信息。

[Background and objectives]Tobacco leaf part recognition is of great significance to the formulation design and quality control of cigarette products.Near infrared spectroscopy(NIR)can be used to realize fast and nondestructive online identification of tobacco leaf parts.Aiming at difficulties in spectral feature extraction of tobacco leaves,a NIR-BYOL tobacco leaf part recognition method is proposed by using the BYOL model with strong feature extraction in this study.[Methods]Data enhancement was realized by differential spectrum fusion,and BYOL's online network and target network were realized by using convolutional self-coder and multi-layer perceptron.By taking the mean square error of the output of the online network and target network as the loss function,the extracted features were classified by SVM to recognize the tobacco leaf part information through the minimum loss optimized coding value.The effects of factors such as data enhancement methods,convolution kernel size and activation function on the model were comparatively analyzed.[Results]The combination of first order differential fusion and second order differential fusion was found to be the best data enhancement method.The best parameter of the contrast learning model was convolution kernel 11*1,and the activation function was ELU.The average recognition rate of the model was 91.79%.Compared with SVM,PCA+SVM and PLS-DA methods,the accuracy of NIR-BYOL model has been significantly improved by 13.12%,15.79%and 16.79%,respectively.[Conclusion]Combining near infrared spectroscopy with comparative learning model can effectively classify and recognize the tobacco leaf part information.