卷积神经网络(CNN)在图像分类识别领域应用广泛, 但其在近红外光谱分类中的研究还未见报道, 对基于CNN的近红外光谱分类建模方法进行了研究。 针对近红外光谱数据的特点, 提出了一种改进的卷积神经网络建模方法, 对CNN经典模型LeNet-5所做改进: ①将方形矩阵卷积核改为适用于一维近红外光谱的向量卷积核; ②简化网络结构, 将LeNet-5结构中C5, F6及输出层改为单层感知机。 同时, 采用隔点采样的方法对近红外光谱降维, 加快收敛速度; 并对卷积核尺寸对建模结果的影响进行了研究。 以我国东北、 黄淮、 西南三大烤烟产区的600个中部烟叶样本的近红外光谱为实验对象, 建立烟叶产区分类NIR-CNN模型。 该模型对训练集和测试集的判别准确率为98.2%和95%。 实验结果表明, 应用卷积神经网络可对近红外光谱数据准确、 可靠地判别分类; 烟叶产区NIR-CNN建模方法可为卷烟企业烟叶原料科学合理利用提供指导, 为维护卷烟产品的质量稳定有重要意义; 基于卷积神经网络的近红外光谱判别方法也可推广到其他农产品的分类应用中。

Convolutional neural network (CNN) was widely used in image classification and recognition but its application in near infrared spectroscopy has not been reported. Therefore, the near-infrared spectroscopy classification modeling method based on CNN was studied in this paper. Taking into account the characteristics of near-infrared spectral data, an improved CNN modeling method was presented in this paper, which improves the CNN classical model Lenet-5: ①The square matrix convolution kernel was transformed into a vector convolution kernel for one-dimensional near-infrared spectroscopy. ②The C5, F6 and output layers of the lenet-5 structure were changed to single-layer sensing machines in order to simplify the network structure. At the same time, the method of sampling points was used to reduce the dimensionality of near infrared spectrum and speed up the convergence rate. The influence of convolution kernel size on modeling results was also studied in this paper. NIR-CNN model was established by the near-infrared spectroscopy of 600 central tobacco samples from northeast, Huanghuai and southwest China. The accuracy of the model was 98.2% and 95% for the training set and test set. The experimental results showed that the application of CNN could accurately and reliably identify the near infrared spectrum data. This method provided guidance for the scientific and rational utilization of raw materials of tobacco enterprises, and it was important to maintain the quality stability of cigarette products. The method of near infrared spectroscopy based on CNN could also be applied in the classification of other agricultural products.