AUTHOR=Wen Jun , He Jing TITLE=Agricultural development driven by the digital economy: improved EfficientNet vegetable quality grading JOURNAL=Frontiers in Sustainable Food Systems VOLUME=Volume 8 - 2024 YEAR=2024 URL=https://www.frontiersin.org/journals/sustainable-food-systems/articles/10.3389/fsufs.2024.1310042 DOI=10.3389/fsufs.2024.1310042 ISSN=2571-581X ABSTRACT=In order to solve the issues attributable to the traditional manual vegetable grading method, this paper proposes a deep learning method to grade vegetable quality. In the first place, a vegetable dataset was constructed to solve the current lack of vegetable datasets; the researchers collected images of vegetables, including lettuce, broccoli, tomatoes, garlic, bitter melon, and Chinese cabbage, constituting the original dataset of 3,600 images in total. Furthermore, this paper suggests an enhanced CA-EfficientNet-CBAM model for the task of vegetable quality grading. In this model, the researchers replaced the attention mechanism SE (squeeze-and-excitation) module in the MBConv (MobileNet convolution) structure of the EfficientNet model with the CA module. This modification enables the network to concurrently preserve both the long-term dependencies of features and precise location information. Additionally, a CBAM (channel and spatial attention module), a lightweight attention module, was embedded prior to the last layer of the network to speed up the training of the model and make the network place greater emphasis on nuanced features, further enhancing the accuracy of the analysis of comparable species. Ultimately, the researchers employed the improved EfficientNet model and other comparison models for training and conducted ablation comparison experiments. Moreover, the experimental results demonstrate that the method proposed in this paper achieved the highest classification accuracy on the cabbage vegetable image test set, reaching 95.12%. Furthermore, in comparison with other models, including VGGNet16, ResNet50, and DenseNet169, the accuracy of the improved EfficientNet model on the test set increased by 8.34%, 7%, and 4.29%, respectively. Likewise, the method proposed in this paper also effectively diminishes the model's parameter count.