基于深度卷积条件生成对抗网络的虫害分类算法研究  

Generating pest identification classification against network based on deep convolution conditions

作　　者：王铎 温长吉[1] 王希龙 李卓识[1,3] 于合龙 Wang Duo;Wen Changji;Wang Xilong;Li Zhuoshi;Yu Helong(College of Information Technology,Jilin Agricultural University,Changchun,130118,China;School of Computer and Software,Nanjing University of Information Science&Technology,Nanjing,210044,China;Key Laboratory of Bionic Engineering,Ministry of Education,Jilin University,Changchun,130022,China)

机构地区：[1]吉林农业大学信息技术学院,长春市130118 [2]南京信息工程大学计算机与软件学院,南京市210044 [3]吉林大学工程仿生教育部重点实验室,长春市130022

出　　处：《中国农机化学报》2020年第6期179-187,共9页Journal of Chinese Agricultural Mechanization

基　　金：国家自然科学基金重点项目(U19A2061);国家自然科学基金面上项目(11372155、61472161);国家重点研发技术专项(2017YFD0502001);吉林省自然科学基金(20180101041JC);吉林省教育厅科研规划重点课题(2016186、JJKH20180659KJ)。

摘　　要：虫害识别是防治的关键环节,但是由于虫害自身的动态性、稳定性、变态性以及种群规模庞大形态各异等特点,因此传统图像分类识别方法面临巨大挑战。本文提出一种基于深度卷积条件生成对抗网络的虫害识别方法。首先,利用空间金字塔层(SPP)嵌入鉴别器的卷积层之后,使得分类的图像可以有固定大小,尺寸一致的效果,不仅可以使得模型获取到必要的图像特征,而且大大提高了模型的收敛速度。其次将深度卷积条件生成对抗网络中的交叉熵损失函数改为最小二乘损失函数,使得最小二乘损失在一点达到饱和,从而增强模型训练的稳定性。最后使用RMSprop优化器对模型进行迭代优化,使得模型能够更快地收敛,增强模型稳定性,减小网络整体的波动。统计训练过程中各项参数变化分析,与其他方法相比,深度卷积条件生成对抗网络在分类准确率和实验稳定性方面都有较为明显的优势,该文所提出的分类准确率高达96.8%,明显优于对比算法。试验结果表明该文提出的虫害分为算法具有较高的准确率和较好的稳定性,可以为相关研究工作提供借鉴意义。Pest identification is a key link in prevention and control.However,due to the dynamic nature,concealment,metamorphosis and the large size of the population,the traditional image classification and identification methods face enormous challenges.This paper proposes a pest classification method based on deep convolution conditions to generate an anti-network.Firstly,after using the spatial pyramid layer(SPP)to embed the convolution layer of the discriminator,the classified image can have a fixed size and uniform size,which not only enables the model to acquire the necessary image features,but also greatly improves the convergence of the model speed.Secondly,the cross-entropy loss function in the deep convolution condition generation against the network is changed to the least squares loss function,so that the least squares loss is saturated at one point,thereby enhancing the stability of the model training.Finally,the RMSprop optimizer is used to iteratively optimize the simulation,so that the model can converge faster,enhance the stability of the model,and reduce the overall fluctuation of the network.Compared with other methods,the deep convolution condition generation confrontation network has obvious advantages in classification accuracy and experimental stability.The classification accuracy rate proposed in this paper reaches 96.8%,significantly better than the comparison algorithm.The experimental results show that the proposed pest classification algorithm has higher accuracy and better stability,which can provide reference for related research work.

关 键 词：图像处理 深度卷积 对抗网络 虫害分类 算法优化 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构] S126[自动化与计算机技术—计算机科学与技术]