基于深度学习的农田害虫识别研究  

Research on Farmland Pest Identification Based on Depth Learning

作　　者：马鑫鑫 张巧雨 马越 孙绪程 陈浩 MA Xinxin;ZHANG Qiaoyu;MA Yue;SUN Xucheng;CHEN Hao(School of Computer and Information Engineering,Nantong University of Technology,Nantong Jiangsu 226001,China)

机构地区：[1]南通理工学院计算机与信息工程学院,江苏南通226001

出　　处：《信息与电脑》2022年第24期180-182,共3页Information & Computer

基　　金：江苏省大学生创新创业训练计划项目“基于深度学习的农田害虫精准识别研究”(项目编号:202212056049Y)。

摘　　要：农田害虫降低了农作物的产量和质量,如何有效区分和治理农田害虫成为首要解决的问题。文章紧抓农田环境需求和农民对农作物的产量需求不匹配的痛点,基于卷积神经网络技术识别农田害虫,为农业提供有效的识别方式。采用MobileNetV1、残差神经网络(Residual Network,ResNet)50、视觉几何群网络(Visual Geometry Group Network,VGG)16以及微调预训练模型VGG16共4种网络模型二分类农田害虫图片集。由于样本数据量较少,为防止出现过拟合,使用了数据增强技术,即通过现有训练图片生成更多的训练图片,从而提高泛化能力。实验表明,4种网络模型的准确率分别为88.63%、91.73%、86.49%和90.13%,在农田害虫识别中均具有较好的实际应用效果。The agricultural pests have reduced the yield and quality of crops. How to effectively distinguish and control agricultural pests has become the first problem to be solved. This paper focuses on the pain point of the mismatch between farmland environmental demand and farmers’ demand for crop yield, and realizes farmland pest identification based on convolutional neural network technology,aiming to provide an effective identification method for agriculture. In this paper, four network models, MobileNetV1, Residual Network(ResNet)50, Visual Geometry Group Network(VGG)16, and fine-tuning VGG16, are used to classify farmland pest image sets. Due to the small amount of sample data, in order to prevent over fitting, this paper uses data enhancement, that is, to generate more training pictures from existing training pictures, so as to improve the generalization ability. The experiments show that the accuracy of the four network models are 88.63%, 91.73%, 86.49% and 90.13% respectively, which have good practical application effect in the identification of agricultural pests.

关 键 词：MobileNetV1 视觉几何群网络(VGG)16 残差神经网络(ResNet)50 过拟合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]