基于改进ResNet的马铃薯黑心病近红外光谱检测方法

Detection on Potato Black Heart Disease by Near Infrared Spectroscopy Based on Improved ResNet

摘要:

马铃薯在存储过程中，极易产生黑心病等内部缺陷，严重影响市场价值和食品安全。探索深度学习用于挖掘马铃薯黑心病光谱数据深层特征，将近红外光谱数据二维化，基于残差神经网络（Residual neural network，ResNet），引入卷积注意力模块（Convolutional block attention module，CBAM)增强特征，加入阈值处理模块去除噪声，实现了马铃薯黑心病的快速无损检测。探索适用于马铃薯黑心病检测的光谱二维化方法，通过对比格拉米角场（Gramian angular field，GAF）、马尔可夫转移场（Markov transition field，MTF）、递归图（Recurrence plot，RP）和波长顺序转换4种方法，发现GAF、MTF和RP这3种方法与波长顺序转换相比效果更好，经过MTF处理后建模效果最佳，训练集准确率达到99.60%。通过比较不同模型性能，发现改进ResNet模型测试集准确率为9765%，比偏最小二乘判别分析（Partial least squares discriminant analysis，PLS-DA）、支持向量机（Support vector machines，SVM）、MobileNet、ResNet分别提高5.89、7.07、3.53、2.36个百分点，MobileNet、ResNet和改进ResNet神经网络模型建模效果优于传统化学计量学方法PLS-DA和SVM。

Abstract:

Potatoes are highly susceptible to internal defects such as black heart disease during storage, which seriously affects market value and food safety. To explore the problem of deep learning in mining the deep features of potato black heart disease spectral data, the near-infrared spectral data were two-dimensionalized, based on residual neural network (ResNet), convolutional block attention module (CBAM) was introduced to enhance the features, and a threshold processing module was added to remove the noise. The features were enhanced by introducing the CBAM, and the noise was removed by adding the thresholding module, which realized the rapid and nondestructive detection of black heart disease in potato. To explore the spectral two-dimensionalization method applicable to the detection of potato black heart disease, four methods, namely, Gramian angular field (GAF), Markov transition field (MTF), recurrence plot (RP) and wavelength-order conversion, were compared and analyzed. It was found that the three methods GAF, MTF and RP were better compared with wavelength-order transformation, and the best modeling effect was achieved after MTF processing, and the accuracy of the training set reached 99.60%. By comparing the performance differences of different models, it was found that the test set accuracy of the improved ResNet model was 97.65%, which was better than that of partial least squares discriminant analysis (PLS-DA), support vector machines (SVM), MobileNet and ResNet by 5.89, 7.07, 3.53 and 2.36 percentage points, respectively, and the traditional chemometrics methods PLS-DA and SVM were not as effective as neural network models such as MobileNet and ResNet in modeling.