基于卷积神经网络的家蚕病害识别研究  

Research on recognition of silkworm diseases based on Convolutional Neural Network

作　　者：石洪康 肖文福[1] 黄亮[2] 胡丛武 胡光荣[1] 张剑飞[1] Shi Hongkang;Xiao Wenfu;Huang Liang;Hu Congwu;Hu Guangrong;Zhang Jianfei(Sericultural Research Institute,Sichuan Academy of Agricultural Sciences,Nanchong,637000,China;State Key Laboratory of Silkworm Genome Biology,Southwest University,Chongqing,400700,China)

机构地区：[1]四川省农业科学院蚕业研究所,四川南充637000 [2]西南大学家蚕基因组生物学国家重点实验室,重庆市400700

出　　处：《中国农机化学报》2022年第1期150-157,共8页Journal of Chinese Agricultural Mechanization

基　　金：智能农业装备研发南充市重点实验室项目(NCKL202008);南充市研发资金(20YFZJ0042);现代农业产业技术体系专项(CARS-18-ZJ0401);现代农业学科建设项目(2021XKJS094)。

摘　　要：病害是我国养蚕业健康发展面临的主要威胁之一,为研究机械化养蚕模式下的家蚕病害防治方法,采用卷积神经网络进行家蚕病害图像的识别研究。首先在实际环境下,采用饲养和添食病原的方法,集中获取家蚕品种芳·秀×白·春在大蚕期的部分生长阶段下患脓病、微粒子病、白僵病、细菌病、农药中毒以及健康状态的样本,并开展图像采集工作,构建出家蚕病害图像数据集。其次采用特征融合和缩减结构的方法,对残差神经网络进行部分改进,以避免直接使用该算法会导致不必要的计算耗损。最后进行家蚕病害识别试验。结果表明:卷积神经网络能够高效准确识别家蚕病害图像,使用改进的算法在测试集上的准确率达到94.31%,与标准的残差神经网络准确率相当,但训练的参数量仅为原来的1/3,且识别效率大幅提升,更有利于网络的训练与部署。Diseases are one of the main threats facing the healthy development of the sericulture industry in China.In order to study methods of prevention and control of silkworm disease under mechanized sericulture mode,a convolutional neural network is used to carry out recognition of silkworm disease images in this paper.First of all,by breading and feeding pathogens,healthy samples and samples of silkworm species Fang·Xiu·Bai·Chun with nuclear polyhedrosis disease,nosema bombycis,beauveria bassiana,bacterial disease,and pesticide poisoning in the stage of larva were obtained.Images collection was conducted to construct a silkworm disease image dataset.Secondly,the method of feature fusion and structure reduction was adopted to modify the residual neural network to avoid unnecessary calculation loss caused by the direct use of the algorithm.Finally,a silkworm disease recognition test was performed.The results show that the convolutional neural network can efficiently and accurately recognize silkworm disease images.The accuracy rate of the modified algorithm on the test set reaches 94.31%,which is equivalent to the standard residual neural network accuracy rate.However,the amount of training parameters is only 1/3 of the original network,which is more conducive to training and deployment of the network,which has a wide range of application prospects.

关 键 词：家蚕 病害识别 卷积神经网络 深度学习 

分 类 号：S887.3[农业科学—特种经济动物饲养]