基于改进型EfficientNet模型的木薯病害识别方法研究  

Research on Cassava Disease Recognition Methods Based on Improved EfficientNet Model

作　　者：曾姣艳 林思涛 谢亚君 曾美艳[3] ZENG Jiaoyan;LIN Sitao;XIE Yajun;ZENG Meiyan(Big Data Institute,Fuzhou University of International Studies and Trade,Fuzhou 350202,China;Key Laboratory of Data Science and Intelligent Computing,Fuzhou 350202,China;School of Commerce&Tourism,Chenzhou Vocational Technical College,Chenzhou Hunan 423000,China)

机构地区：[1]福州外语外贸学院大数据学院,福州350202 [2]数据科学与智能计算重点实验室,福州350202 [3]郴州职业技术学院商贸旅游学院,湖南郴州423000

出　　处：《西南大学学报（自然科学版）》2024年第6期197-208,共12页Journal of Southwest University(Natural Science Edition)

基　　金：国家自然科学基金项目(12371378);福建省自然科学基金项目(2022J01378).

摘　　要：为实现木薯病害图像的快速、准确识别,提出一种基于EfficientNet模型的木薯病害识别方法.首先针对输入样本的分布不平衡问题,通过Mixup、CutMix及GridMask这3种数据增强方法对数据进行增强,数据增强后由EfficientNet-B4模型提取特征,然后引入warmup结合余弦退火优化学习率防止模型在初期发生过拟合及后期收敛速度慢的情况.实验结果表明,所采用模型相较于近年来主流的VGG16及ResNet101模型不仅参数量远小于两者,在木薯病害图像分类上的表现也优于两者,且其计算量更少,模型精度更高,训练速度更快,符合实际应用的要求.EfficientNet模型在木薯病害数据上的分类准确率可达90%.An EfficientNet model based cassava disease recognition method was proposed for fast and accurate recognition of cassava disease images.The method first addressed the imbalance in the distribution of the input samples,performed the data enhancement by three data enhancement methods of Mixup,CutMix and GridMask.The feature was extracted by EfficientNet-B4 model after data enhancement,then warmup combined with cosine annealing learning rate was introduced to prevent the model from overfitting at early stage and slow convergence speed at late stage.The experimental results show that compared with the mainstream VGG16 and ResNet101 models in recent years,the model used in this paper not only has a much smaller number of parameters than the two models,but also has a better performance on classification of cassava leaf disease image than that of two models.It is less computationally intensive,with higher model accuracy and faster training,which meets the requirements of practical applications.EfficientNet model can achieve 90%classification accuracy on cassava leaf disease data.

关 键 词：木薯病害图像 数据增强 EfficientNet模型 余弦退火 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]