摘要
针对病虫害症状相似导致类间差异小、难以区分的问题,提出一种基于Res2NeXt50改进模型的农作物病虫害检测算法。首先,在Res2Net50模型中进行分组卷积得到Res2NeXt50模型,提高了模型在细粒度层面的特征提取能力。然后,将7×7卷积换成新的混合卷积,提取局部和全局特征;使用高斯误差线性单元(Gaussian Error Linear Unit,GELU)函数代替残差块中的修正线性单元(Rectified Linear Unit,ReLU)函数,提高鲁棒性;改进下采样来增强信息流通性;调整网络层数,以减少模型计算量。其次,在训练中使用标签平滑(Label Smoothing)和指数移动平均(Exponential Moving Average,EMA)来提高模型的泛化能力。在重组的AI Challenger 2018农作物病虫害数据集上进行实验,结果表明改进模型的准确率高达98.79%,参数量为18.20M,FLOPs为3.73G。同时,该模型在Plantvillage和Plant_leaves数据集中分别达到了99.89%和99.23%的准确率。所提出的算法模型识别准确率高,泛化能力强,更符合实际应用需求。
Aiming at the problem of small differences and indistinguishability between classes due to similar symptoms of disease and insect pest,a crop disease and insect pest detection algorithm based on the improved Res2NeXt50 model is proposed.Firstly,grouping convolution is performed in the Res2Net50 model to obtain the Res2NeXt50 model,which improves the feature extraction capability of the model at the fine-grained level.Then,the 7×7 convolution is replaced with a new mixed convolution to extract local and global features.The Gaussian Error Linear Unit(GELU)is used to replace the Rectified Linear Unit(ReLU)in the residual block in order to improve robustness.Downsampling is improved to enhance information flow,and the number of network layers is adjusted to reduce the amount of model computation.Secondly,Label Smoothing and Exponential Moving Average are used in training to improve the generalization ability of the model.Experiments were carried out on the restructured AI Challenger 2018 crop disease and insect pest dataset,which showed that the accuracy of the improved model was as high as 98.79%,the number of parameters was 18.20M,and the FLOPs was 3.73G.Meanwhile,the model achieves 99.89%and 99.23%accuracy in Plantvillage and Plant_leaves datasets,respectively.The proposed algorithm model has high recognition accuracy and strong generalization ability,which is more in line with practical application requirements.
作者
白雪松
吴建平
景文超
崔亚楠
康小霖
BAI Xue-song;WU Jian-ping;JING Wen-chao;CUI Ya-nan;KANG Xiao-lin(School of Information Science&Engineering,Yunnan University,Kunming 650504,China;Yunnan Provincial Electronic Computing Center,Kunming 650223,China;Digital Media Technology Key Laboratory of Universities and Colleges in Yunnan Province,Kunming 650223,China)
出处
《计算机技术与发展》
2023年第5期145-151,共7页
Computer Technology and Development
基金
云南省重大科技专项计划项目(202002AD080001)
云南大学第一届专业学位研究生实践创新项目(2021Y183)。
