Abstract
For image classification of crops, most convolutional neural network (CNN) models have low accuracy, especially in modern agricultural environments. Furthermore, crop disease images create more difficulties for classification owing to the morphological and physiological changes of organs, tissues, and cells. Here, we propose a CNN model named CAMFFNet (coordinate attention-based multiple feature fusion network) for tobacco disease identification under field conditions. The CAMFFNet model has three multiple feature fusion (MFF) modules. Each module is composed of two residual blocks. The MFF module is concatenated by max-pooling downsampling layers at different locations in the residual blocks to realize a fusion between features of multiple depths, thereby reducing the loss of tobacco disease information. Furthermore, to enhance the ability to extract effective feature information of tobacco diseases and to alleviate the impact of the field environment, coordinate attention (CA) modules are included between each multiple feature fusion module. The obtained results show that the CAMFFNet model achieved an accuracy of 89.71 % on the tobacco disease test set. The accuracy was 3.36 %, 4.7 %, 4.7 %, 2.91 %, 8.05 %, 4.92 %, 10.07 %, and 2.91 % higher than those of the classic CNN models VGG16, GoogLeNet, DenseNet121, ResNet34, MobbileNetV2, MobbileNetV3 Large, ShuffleNetV2 1.0×, and EfficientNetV2 Small, respectively. In addition, the CAMFFNet model’s number of parameters is only 2.37 million. The results demonstrate that the CAMFFNet model has a high potential for tobacco disease recognition in mobile and embedded devices.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have