Application of Hyperspectral Imaging and Deep Convolutional Neural Network for Freezing Damage Identification on...

In this paper, the feasibility of identifying freezing damage on the endosperm side and embryo side of single corn seeds was studied by combining hyperspectral imaging technology and the deep convolutional neural network (DCNN) method. Firstly, hyperspectral image data of the endosperm and embryo side of three freezing-damage categories of corn seeds were collected, and the average spectra of the endosperm part and embryo part were obtained with the range of 450-979 nm. After the spectral data were pre-processed by non-pretreatment or standard normal variation (SNV) pretreatment, a support vector machine (SVM) and a DCNN model were established for freezing-damage identification. Finally, multiple evaluation indexes (including accuracy, sensitivity, specificity, and precision) were used to comprehensively evaluate the performance of the SVM and DCNN models in the whole waveband. The results showed that the DCNN model obtained better performance in accuracy, sensitivity, specificity, and accuracy. The values of each category, especially for the category-2 and category-3 testing sets of the SVM model, were lower than those of the DCNN model. The classification results of the embryo-side corn seeds were better than those of the endosperm side. The accuracy value of the testing set of the DCNN model on the embryo side was higher than 96.7%, while the accuracy value of the DCNN model on the endosperm side was lower than 93.8%. The specificity values of the SVM and DCNN models were both higher than 94%. In addition, the sensitivity and precision values of the category-2 testing set of the embryo-side DCNN model increased by at least 2.8% and 4.8%. The sensitivity value of the category-3 testing set of the DCNN model was improved by at least 8.2% and 4.4%. These results of the embryo side of the corn seed showed significant improvement in the training and testing set. This study proved that the DCNN model can accurately and quickly identify single freezing-damage corn seeds, which provided a theoretical basis for constructing an end-to-end recognition and classification model of frozen corn seeds.