基于回归CNN的烟叶近红外光谱模型研究

Study on Near Infrared Spectroscopy Model of Tobacco Leaves Based on Regression CNN

为了更大程度提取近红外光谱中的深层次关键特征,论文应用改进的回归式卷积神经网络算法,将传统的卷积神经网络架构中池化层移除,最顶层的线性分类层用回归层进行替代,构建卷积神经网络回归(CNNR)模型。为了验证该算法的有效性,论文经多次实验、对比评价指标,筛选出最佳模型:总糖、总烟碱和氯离子最佳CNNR模型的相关系数R分别为0.9318,0.941,0.933,交叉验证的RMSECV分别为0.7052,0.0710,0.0971。实验结果表明:CNNR模型抽提的特征光谱数据对三个指标有很强的解释能力,对烟叶化学成分的有较好的预测性能和综合表达能力。

In order to extract the deep critical features of near-infrared spectroscopy to a greater extent,a novel predictionmethod,namely,regression-based convolutional neural network(CNNR),which removes the pooling layer and substitute the re-gression layer for the linear softmax classification layer at the top of the general CNN’s structure,has been proposed to develop thequantitative model for the tobacco constituents. In order to verify the effectiveness of the CNNR algorithm,the models for total sug-ar,total nicotine and chlorine in tobacco are built for which correlation coefficient R are 0.9318,0.941,0.933,respectively andRMSECV of cross validation were 0.7052,0.0710,0.0971,respectively. The results indicate that the extracted features have astrong ability to interpret the spectral data and have better prediction performance and comprehensive expression ability of tobaccochemical components.