基于EfficientNet的木薯叶病变自动分类模型

Automatic classification model for cassava leaf disease based on EfficientNet

针对数据集中标注存在错误的情况下,传统的分类方法的模型在学习错误特征时过于自信导致准确率低的问题,提出了基于EfficientNet的自动分类模型。首先,对输入图像做数据增强,扩充数据后由EfficientNet提取特征;然后,引入标签平滑和随机丢弃节点,使模型不会过于自信,提高模型的泛化能力;最后采用双稳态逻辑损失进行预测。训练中通过对数据集做分层交叉验证来避免过拟合。实验结果表明,所提模型不仅比参数量低的模型表现更加优异,甚至比几倍于自身参数量的如ResNeXt或者十几倍于自身参数量的如RepVGG等模型也有更好的表现。所提模型计算量更小,推理速度更快,算法精度更高,更符合实际落地的要求。所提模型在木薯叶病变公共数据集上的准确率达到了89.66%。

In order to solve the problem that traditional classification method was too confident and the accuracy was too low in learning the wrong features when there were some errors in the annotation,an automatic classification model based on EfficientNet was proposed. Firstly,data augmentation was performed on the input images,after the dataset was expanded, features were extracted by EfficientNet. Then,label smoothing and dropout were introduced to prevent the model from overconfidence and improve the generalization ability of the model. Finally,bi-tempered logistic loss was used to predict. During training,over-fitting was avoided by doing stratified k-fold on the dataset. Experimental results show that the proposed model not only performs better than models with lower parameter amount,but also performs better than models with several times its own parameter amount,such as ResNeXt,or models with more than ten times its own parameter amount, such as RepVGG. The proposed model has smaller calculations,faster inference speed,higher accuracy,and is more in line with landing requirements. The proposed model has an accuracy of 89. 66% on the public dataset of cassava leaf diseases.