Non-Destructive Identification Method of Famous Rice Based on Image and Spectral Features of Hyperspectral Imaging with Convolutional Neural Network
Weng Shi-zhuang,Tang Pei-pei,Zhang Xue-yan,Xu Chao,Zheng Ling,Huang Lin-sheng,Zhao Jin-ling
DOI: https://doi.org/10.3964/j.issn.1000-0593(2020)09-2826-08
2020-01-01
Spectroscopy and spectral analysis
Abstract:High-quality rice contains more nutritional value and higher economic value. In order to earn more benefits, some unscrupulous merchants have adulterated high-quality rice or even replaced it with low-quality rice, which has harmed consumer interests and rice trade, and has hurt producers Production motivation. This paper hopes to develop a method for non-destructive identification of high-quality rice based on the features of images and spectra of hyperspectral imaging and deep learning networks. First, hyperspectral images in the 400 similar to 1000 nm range of seven representative rice varieties in China were collected, and the spectra, texture, and shape features of each type of rice were extracted. The spectral features were pre-processed using the multiple scattering correction algorithms to eliminate spectral scattering. Successive projections algorithm (SPA) competitive adaptive weighting algorithm (CARS) and their cascade method (CARS-SPA) were used to select important wavelengths of spectral features. Important variables of shape and texture features were selected using SPA. Finally, convolutional neural network (CNN) was applied to fuse the above-mentioned various features to build rice varieties recognition model, while K-Nearest Neighbors (KNN) and Random Forest (RF) were used for comparison and analysis. The experimental results showed that the classification accuracy of the model constructed using the full spectroscopy reached more than 80%. Among them, KNN had the worst modeling effect and RF had a better effect. In particular, the performance of the CNN model was the best, with training set classification accuracy (ACC(T)) of 92.96% and prediction set classification accuracy (ACC(p)) of 89.71%. Compared with the full spectroscopy, the spectroscopy of the important wavelengths had worse classification accuracy. In order to further improve the accuracy of rice varieties identification, texture and shape were combined with spectral features, and the optimal result came from the model constructed of important variables of shape and spectroscopy. Among them, ACC(T) and ACC(p) of KNN were 69% and 67%, respectively. The RF model accuracy corresponded to ACC(T) = 99.98% and ACC(p) = 89.10%. The CNN model performed best with ACC(T) and ACC(p) of 97.19% and 94.55%. In addition, the classification effect of spectroscopy and texture fusion was worse than using only spectroscopy, indicating that texture features weakened the classification result. For classification models, the performance of CNN was significantly better than the two machine learning methods, which could provide better classification results. All in all, the important variables of shape and spectroscopy combined with CNN models could accurately identify high-quality rice varieties. The proposed method can also be applied to the identification of the variety, attribution and grade of other agricultural products.