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

作　　者：姜天宇 赵晓林[1] 赵搏欣 李伟龙[1] 吴梦瑶 JIANG Tianyu;ZHAO Xiaolin;ZHAO Boxin;LI Weilong;WU Mengyao(Equipment Management and Unmanned Aerial Vehicle Engineering College,Air Force Engineering University,Xi’an Shaanxi 710000,China)

机构地区：[1]空军工程大学装备管理与无人机工程学院,西安710000

出　　处：《计算机应用》2022年第S01期64-70,共7页journal of Computer Applications

基　　金：国家自然科学基金资助项目(GKJJ0217060301);陕西省自然科学基金资助项目(2021JQ⁃354)。

摘　　要：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.

关 键 词：木薯叶病变分类 标签噪声 标签平滑 双稳态逻辑损失 随机丢弃节点 EfficientNet 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]