Abstract:

To improve the accuracy of identifying the maturity of tobacco leaves at different harvesting stages， this study systematically collected images of under-ripe， ripe， and over-ripe tobacco leaves from the upper， middle， and lower parts of the plant. We proposed an improved VGG-16-based deep learning method for tobacco leaf maturity recognition. This method leverages a pre-trained VGG-16 model as its foundation， incorporating transfer learning， fine-tuning， and convolutional layer feature fusion to enhance the model's performance in tobacco leaf maturity recognition tasks. Freezing the convolutional and subsequent layers， adding BN layers， and using the Adam optimizer further improved training efficiency， avoided over fitting， and enhanced the model's robustness and accuracy. Experimental results showed that the improved VGG-16 model had a high accuracy advantage in tobacco leaf maturity recognition tasks， with a test set accuracy of 99.7%， surpassing classical machine learning methods such as BP neural networks， support vector machines， original VGG-16 and VGG-19， AlexNet， and ResNet50. The model comprised 14 721 353 parameters， a model size of 58.9 M， and a single image recognition time of 0.024 9 seconds， demonstrating the advantages of low computational and storage resource requirements and rapid recognition speed. Further visual analysis of the recognition results of tobacco leaf images with different maturities using the Score-CAM algorithm revealed that the central region of the main vein of tobacco leaves served as a primary differentiating feature for tobacco leaves with different maturities across different stalk positions， providing key information for the recognition model， there by clarifying the chemical substance transformation patterns occurring during the maturity process of tobacco leaves from distinct plant sections. The improved VGG-16 deep learning model proposed in this study exhibits high accuracy and efficiency in identifying the maturity of tobacco leaves at different harvesting stages. It is poised to provide precise and effective decision support for tobacco harvesting and production. Future research will focus on exploring alternative feature fusion strategies and network structures to further improve the generalization ability and robustness of tobacco leaf maturity recognition across different production areas and years.

Key words: tobacco maturity, deep learning, improved VGG-16, transfer learning, image classification