Abstract:
Aiming to achieve intelligent recognition of flue-cured tobacco leaves grade quickly and accurately, the images of front and back tobacco leaves were taken by mobile phone, and a new model (VGG16-Dense) integrated with VGG16 and DenseNet was constructed. The validity of the model was verified by using twenty-four types of front and back leaf images of cv. Cuibi-1 and Yunyan87. The model was also compared with five other network models, i.e. DenseNet121, ResNet50, AlexNet, VGG16 and GoogLeNet. The results shows that excellent values appeared in all evaluation indicators (accuracy, precision, recall, F1-score and avg-loss) of validation set for VGG16-Dense, and the evaluation indicators of test set for VGG16-Dense performed optimal compared with that of other network models. VGG16-Dense exhibited superior generalization ability and fewer misjudgments, with the accuracy, precision recall, F1-score, and avg-loss reaching 92.71%, 93.07%, 92.71%, 92.72%, and 0.22, respectively. VGG16-DENSE network model can intelligently distinguish the grade, the front and back, and the species for flue-cured tobacco leaves at the same time. This provides a theoretical guidance for the intelligentized grading of primary flue-cured tobacco acquisition.