【作者】 刘宁；

【导师】 吕辰刚； 焦强；

【作者基本信息】 天津大学 ， 电子与通信工程， 2019， 硕士

【摘要】 我国作为农业大国,农业的发展影响着国内民生和国家经济发展。农作物病虫害的识别是农业丰收的第一道保障,而传统的人工识别和专家系统因为准确率和响应速度等因素,不符合农业现代化的要求。深度学习是人工智能发展的重要分支和实现方式,被应用到了图像识别、机器翻译等多个领域。通过深度学习对农作物病虫害进行检测和分类是农业现代化的发展趋势和要求。本文基于Ai Challenger发布的农作物病虫害数据集,使用神经网络对61类病虫害数据进行了分类。首先使用图像变化和互联网搜集的病虫害图片对原数据集进行了增广,并使用类似Alex Net的网络结构验证了增广的有效性。之后使用更深层的神经网络对增广后数据进行训练,并对比了不同类别的疾病的分类情况,根据疾病造成的危害严重程度和疾病处理的难度,分析了网络评价的侧重点。由于农作物疾病早期不易发现,导致其样本数量相对疾病发展中后期较少,进而引起了样本间数量不平衡的问题。本文通过数据集抽样与模型融合的方式,改善了这一问题,提高了同一疾病不同发展阶段分类的准确率。由于实际生产活动中,大量有标注病虫害数据不易获得,本文使用生成对抗网络(GANs)对上述数据集中的少量样本采用半监督学习的方式进行训练。使用50%的标注样本,可获得接近100%标注样本的监督学习的准确率。对原GANs静态训练方式进行优化,提出了一种动态反馈调节GANs训练的方法,提高了网络的收敛速度和稳定性。对原GANs的半监督学习损失函数进行改进,降低了原损失函数中生成数据造成的等效数据不均衡程度,提高了网络的分类准确率。更多还原

【Abstract】 As a big agricultural country,the development of agriculture affects the people’s livelihood and national economic development.The identification of plant disease is the first guarantee of agricultural harvest.Traditional artificial identification and expert system do not meet the requirements of agricultural modernization because of the accuracy and respond speed.Deep learning is an important branch of the development of artificial intelligence,and has been applied to many fields such as image recognition,machine translation and so on.Detection and classification of crop diseases and insect pests through deep learning is a hot area in the development of modern agriculture.Based on the dataset published by AI Challenger,61 kinds of pests and diseases data were classified using neural networks.Firstly,the original data set is augmented by transformed images and images collected on the Internet,and the Alex Net-like network is used to verify the effectiveness of the augmentation.Secondly,a deeper neural network is used to train on the augmented data,and the classification of various types of diseases is analyzed.Because crop diseases are not easy to detect in the early stage,the number of samples is relatively small compared with the late-stage ones,which leads to the problem of imbalance in the number of samples.This problem is improved by data set sampling and model fusion,and the classification accuracy of the same disease at different stages of development is improved.Because the labeled data of pests and diseases are not easy to obtain in actual production activities,this paper uses GANs to train on a small number of samples in the above data sets in the way of semi-supervised learning.Compared with the original supervised learning method,it reduces the demand for labeled data.Using 50% labeled samples,we can get the accuracy close to 100% of supervised learning data.A feedback dynamic mediation GAN training method is proposed to improve the convergence speed and stability of the network.The loss function of GAN semi-supervised learning is improved,the degree of data imbalance caused by generated data is reduced,and the accuracy of the network is improved.更多还原