Research on leaf identification of table grape varieties based on deep learning

【Objective】With the continuous development of grape varieties, problems such as variety confusion and inaccurate identification would occur in the actual production and scientific research work, and accurate identification of grape varieties has become more and more difficult It is urgent to explore a nondestructive, efficient and environment-friendly identification method. This work amied to provide reference for the protection, utilization, and classification research of table grape varieties. 【Methods】In this study, the leaf images were taken in the National Grape Germplasm Nursery (Zhengzhou) of Zhengzhou Fruit Tree Institute, Chinese Academy of Agricultural Sciences. The images of mature leaves of the 68 common table grapes were taken under natural condition in the field. The leaves were fully expanded without obvious symptoms of nutrition deficiency, pathogen infection and insect damage. The sampling position was at the 7th-9th nedes of the new shoots. The front images of differ-ent leaves were taken, and a dataset of 29 713 fresh grape leaves was constructed. In the realm of automatic recognition, four distinct convolutional neural network models were deployed: GoogleNet, ResNet-50, ResNet-101, and VGG-16.【Results】Through fair comparison of all convolutional neural networks, under optimal parameters, ResNet-101 performed best in the identification of table grapes, with an accuracy of 97.13%, ResNet-50 was slightly lower, with an accuracy of 97.06%, and VGG-16 and GoogleNet models had an accuracy of less than 95%. When ResNet-101 was used as the classification model, the optimized parameters were the learning rate of 0.005, the minimum batch of 32, and the number of iterations was 50. Under this parameter, the classification performance was the best, and the classification accuracy was as high as 97.99% . The model accuracy and LOSS value of ResNet- 101 model was significantly higher than those of other models. The initial accuracy was the highest, the convergence was faster and more stable, the final accuracy was the highest, the initial LOSS was the lowest, the LOSS decreased faster, and the final LOSS was relatively stable. Among the 68 varieties identified by the ResNet-101 model, the prediction accuracy of the 23 varieties was 100%, and the average recognition accuracy of the 68 varieties reached 94.90%; The prediction accuracy of the ResNet-50 for the 13 varieties was 100%, and the average recognition accuracy of the 68 varieties reached 90.38%; The prediction accuracy of the VGG-16 for the 11 varieties was 100%, and the average recognition accuracy of the 68 varieties was 85.45%; The prediction accuracy of the GoogleNet model was 100% for only 5 varieties, and the average recognition accuracy of the 68 varieties was 78.79%. In contrast, the prediction accuracy of the ResNet- 101 model was significantly higher than that of the ResNet- 50, GoogleNet and VGG- 16 models, and the difference in recognition accuracy between varieties was smaller and more stable. In ResNet-101 model, the Recall rate of the 18 varieties reached 100%, and the average Recall rate of the 68 varieties reached 94.19%; In the ResNet-50 model, the Recall rate of the 6 varieties reached 100%, and the average Recall rate of the 68 varieties reached 88.71%; The average Recall rate of the 68 varieties of VGG-16 was 82.83%; The GoogleNet had only two varieties, Crimson Seedless and sunshine rose, with a Recall rate of 100%, and the average Recall rate of the 68 varieties was 74.44%. In contrast, the ResNet-101 model was significantly better than the ResNet-50 GoogleNet, VGG-16. The Recall rate among the varieties was more stable. The F1 value of creson seedless, Xianfeng and longan varieties in the ResNet-101 model reached 1, and the average F1 value reached 0.94. The difference of the F1 value among the varieties was small, the reliability of the model was high, and the model effect was more stable. The F1 value of VGG-16 model for Crimson Seedless reached 1, and the average F1 value reached 0.82; No F1 value of the ResNet-50 and GoogleNet models reached 1, and their average F1 values were 0.88 and 0.74, respectively. The Grad-CAM algorithm was used to output the weighted gradient heat map in the final accretion layer, and the network model was visualized. The results showed that the four convolutional neural networks could accurately identify the main characteristics of the leaves, and the leaf texture, vein and edge of the leaves had the greatest impact on variety recognition.【Conclusion】The ResNet-101 model had the highest overall recognition accuracy, the lowest LOSS value, the higher average recognition accuracy and Recall rate of varieties, and could get a better model with fewer iterations, which would take less time. The Grad-CAM algorithm was used to evaluate the classification effect of four convolutional neural networks, and all of them could accurately identify the main features of the leaves. The rapid and accurate recognition of the table grapes was realized. Therefore, the deep learning network model could complete the automatic real-time recognition of the table grapes.