Discrimination of unsound wheat kernels based on deep convolutional generative adversarial network and near-infrared hyperspectral imaging technology

Abstract

The quality of wheat kernels is critical to ensure crop yields. However, in actual breeding work, unsound kernels are scarce compared to healthy kernels. Limited data sets or unbalanced data sets make it difficult for many algorithms to accurately identify kernels in different states. A novel method based on deep convolutional generative adversarial network (DCGAN) and near-infrared hyperspectral imaging technology was proposed to identify unsound wheat kernels in this paper. Three classifiers, convolutional neural network (CNN), support vector machine (SVM) and decision tree (DT) were used. After expanding the samples, the results showed that the accuracy of the test set of the DT model increased from 51.67% to 80.83%, a total increase of 29.16%. And the CNN and SVM models increased by 8.34% and 14.17% respectively. This demonstrated that the DCGAN method had the ability to generate reliable data samples for unbalanced data sets for improving the performance of the classifier. On this basis, the training samples are further expanded for improving the performance of the classifier. The results showed that CNN model gained the most from incremental data, and its accuracy rate had been continuously improved from 79.17% to 96.67%, a total increase of 17.50%. This also demonstrated that the DCGAN method had the ability to expand a limited data set. In general, the joint model based on DCGAN and CNN combined with hyperspectral imaging technology had a good prospect in the identification of unsound kernels.