基于SimAM-ConvNeXt-FL的茶叶病害小样本分类方法研究

为实现茶叶病害精准分类，针对茶叶病害图像分类中小样本问题及类别分布不均的问题，提出了一种基于迁移学习的SimAM-ConvNeXt-FL模型的病害图像分类方法。首先在ConvNeXt模型中加入SimAM模块，以加强复杂特征的提取。其次针对样本分布不均问题，将Focal Loss函数作为训练过程中的损失函数，通过增加数量较少样本的权重来减小样本分布不均的影响。最后使用SimAM-ConvNeXt-FL模型对Plant Village数据集训练，将训练得到的参数迁移到实测的茶叶病害图像上并进行微调，减少过拟合带来的影响，设置消融实验证明模型改进的有效性，并与不同分类模型(AlexNet、VGG16、ResNet34模型)分别进行对比实验。实验结果表明，SimAM-ConvNeXt-FL模型识别效果最佳，准确率达96.48%, SimAM-ConvNeXt-FL模型较原ConvNeXt模型在茶煤病、茶藻斑病、茶炭疽病、健康叶片和茶白星病的F1值分别提高4.46、3.76、0.43、0.22、5.23个百分点。结果表明本文提出的模型具有较高的分类准确率与较强的泛化性，可推进茶叶病害分类工作发展...

Small Sample Classification of Tea Diseases Based on SimAM-ConvNeXt-FL

In order to realize accurate classification of tea diseases, a disease image classification method based on SimAM-ConvNeXt-FL model of migration learning was proposed to address the small sample problem and uneven distribution of categories in tea disease image classification. Firstly, an SimAM module was added to the ConvNeXt model to enhance the extraction of complex features. Secondly, to address the problem of uneven sample distribution, the Focal Loss function was used as the loss function in the training process, and the effect of uneven sample distribution was reduced by increasing the weights of a smaller number of samples. Finally, the SimAM-ConvNeXt-FL model was used to train the Plant Village dataset, and the parameters obtained from the training were migrated to the measured tea leaf disease images and fine-tuned to reduce the impact of overfitting, and ablation experiments were set up to prove the validity of the model improvement, and comparison experiments were carried out with the different classification models AlexNet, VGG16, and ResNet34 models comparison experiments were conducted respectively. The experimental results showed that the SimAM-ConvNeXt-FL model had the best recognition effect, with an accuracy of 9648%, and the F1 values of the SimAM-ConvNeXt-FL model compared with the original ConvNeXt model for tea coal disease, tea phoma, tea anthracnose, healthy leaves, and tea white star disease were improved by 4.46 percentage points, 3.76 percentage points, 0.43 percentage points, 0.22 percentage points, and 5.23 percentage points respectively. The results showed that the model proposed had high classification accuracy and strong generalizability, which can promote the development of tea disease classification.