文献类型： 外文期刊

第一作者： Chen Dong-ying

作者： Chen Dong-ying;Zhang Hao;Zhang Zi-long;Yu Mu-xin;Chen Dong-ying;Zhang Hao;Chen Lu

作者机构：

关键词： Honeysuckle; Near-infrared spectroscopy; Very-deep (VD); One-dimensional convolutional neural network ( 1D-CNN); Origin traceability

期刊名称：SPECTROSCOPY AND SPECTRAL ANALYSIS （ 影响因子：0.7； 五年影响因子：0.6 ）

ISSN： 1000-0593

年卷期： 2023 年 43 卷 5 期

页码：

收录情况： SCI

摘要： Honeysuckle is an essential medicine for clearing away heat and detoxifying. However, the sources of honeysuckle on the market are complicated, and the most famous honeysuckle produced in Pingyi, Shandong, is often counterfeited. Most existing identification methods are time-consuming, costly, and complex to operate. Therefore, a fast and efficient way to trace the honeysuckles origin is urgently needed. The current one-dimensional convolutional neural network (1D-CNN) identification model based on honeysuckle near-infrared spectroscopy (NIRS) data has the problems of too many parameters and too low model efficiency, high computational complexity and is prone to overfitting. This paper improves the traditional 1D-CNN structure. We use the more efficient VD (Very Deep) structure to replace the hidden layer structure in the conventional 1D-CNN and make adaptive improvements for NIRS data so that the model can be directly applied to one-dimensional NIRS data. The improvement method is divided into three steps: firstly, the design of the feature layer is converted into two constraints optimization design: the first constraint is to set the C value of each convolution layer (the ratio of the size of the convolution kernel and the receptive field) to 1/6, which can improve the efficiency of the network model; the second constraint is to take the size of the top-level sensory domain as the size of the data vector, which can achieve feature extraction of deeper data and reduce overfitting. Secondly, this design minimizes the output feature vector of the feature layer to a smaller size through the downsampling operation. Finally, use two convolutional layers of size 1x5 and a pooling layer with dropout to downsample the data size to a vector of only one vector instead of a fully connected layer for classification, thereby reducing the number of parameters. In the experiment, 500 honeysuckles samples were collected from the main producing areas of Henan, Shandong, Hebei, and Chongqing. The spectral range used in the test is 908 similar to 1 676 nm. The sample set was preprocessed by the KS algorithm, and the training set, validation set and test set were divided by the shuffle algorithm. At last, a honeysuckle origin identification model based on improved 1D-VD-CNN and near-infrared spectroscopy was constructed. The results show that the 1D-VD-CNN training set and test sets accuracy reach 100%, and the loss value converges around 0.001. Compared with the traditional 1D-CNN model, the training set and test set accuracy of the 1D-VD-CNN model are improved by about 0.5% and 1.4%, respectively, and the number of parameters and FLOPs are reduced by nearly 1 M and 20 M, respectively. At the same time, compared with the original spectral data analysis method and the PLS-DA method, it shows that the 1D-VD-CNN model has higher efficiency and better recognition performance for honeysuckle near-infrared spectral classification.

分类号：