【作者】 方敏；

【导师】 胡根生； 丁增辉；

【作者基本信息】 安徽大学 ， 电子与通信工程（专业学位）， 2022， 硕士

【摘要】 中国是全球主要的茶叶生产和消费国,有大面积的茶叶种植园。然而,由于土壤、气候、生态等原因,茶树会受到不同种类的病害感染,从而导致减产。精确识别茶叶病害有利于指导茶农精准喷药,减少因茶叶的病害所造成的损失,保证茶农的经济利益。由于大部分茶树生长在崎岖的山区,植保专家翻山越岭去辨别茶叶病害的效率低成本高。基于图像处理和机器学习的识别方法能够及时且有效识别病害茶叶信息,传统机器学习的方法需要手工提取病害特征,识别精度不高。随着计算机视觉技术的不断发展,基于深度学习的农作物病害自动识别成为热门研究方向,但自然场景下的茶叶病害自动识别存在数据量较少、背景较为复杂等难点问题,本文研究基于深度学习的小样本下茶叶病害识别,主要内容及研究成果如下:1、介绍了本文所使用的茶叶图像相关信息和数据预处理方法,包含采集图像所用设备的型号,地理位置和采集图像的数量。预处理方法分为两步:(1)通过U-Net对茶叶图像进行背景分割,降低复杂背景对识别结果的影响。(2)采用上述茶叶病害图像做为原始训练样本,通过非条件生成模型Sin GAN生成新的训练样本,以进行数据增广。2、采用了一种多卷积神经网络体系结构,即Merge Model,并结合一种新的权值初始化方法用于小样本下的茶叶病害自动识别。多个卷积模块相结合使得Merge Model能够提取到多样的判别特征,因此,Merge Model的识别能力相较单一的神经网络得到了提升。网络中的卷积滤波器权重没有随机初始化,而是将图像病害特征编码到其中,这有助于模型在训练开始时集中学习有用的特征。实验结果表明,Merge Model能将健康茶叶和病害茶叶进行有效区分,并识别出茶白星病,茶叶叶枯病,茶赤叶斑病,茶煤病等茶叶常见病害。与现有方法相比,本文所提方法具有更高的小样本下的茶叶病害识别准确率。3、采用了一种基于图像文本协同表示学习的小样本下的茶叶病害识别算法。由于数据样本较少,仅仅使用图像,所取得的信息有限,难以获得高准确率的识别结果。因此通过添加文本信息作为先验知识,以确保模型在小样本数据集上获得良好的识别结果。本文将图像信息通道和文本增强通道输出内容相结合,利用两种类型信息之间的关联程度和互补性,实现了对病害特征的协同能力识别算法。其中图像信息通道利用视觉Transformer获取到图像模态信息,文本增强通道通过ALBERT结合Text CNN获取到茶叶图像所对应的文本模态信息。实验结果表明,结合两种模态信息的新模型在小样本数据集下能够取得比单独的图像模型或文本模型更好的结果。更多还原

【Abstract】 China is a major tea producing and consuming country in the world,with a large area of tea plantations.However,due to soil,climate,ecology and other reasons,tea trees will be infected by different kinds of diseases,which will lead to the reduction of production.Accurate identification of tea diseases is conducive to guiding tea farmers to spray accurately,reducing the losses caused by tea diseases and ensuring the economic benefits of tea farmers.Because most tea trees grow in rugged mountainous areas,it is inefficient and costly for plant protection experts to identify tea diseases.The identification method based on image processing and machine learning can identify the information of diseased tea timely and effectively.The traditional machine learning method needs to extract the characteristics of diseases manually,and the identification accuracy is not high.With the continuous development of computer vision technology,automatic identification of crop diseases based on deep learning has become a hot research direction.However,there are some difficult problems in automatic identification of tea diseases in natural scenes,such as less data and complicated background.This thesis studies the identification of tea diseases based on deep learning in small samples,and the main contents and research results are as follows:1.Introduce the information and data preprocessing methods of tea images used in this thesis,including the model of the equipment used to collect images,geographical location and the number of images collected.The preprocessing method is divided into two steps:(1)The background of tea image is segmented by U-Net to reduce the influence of complex background on recognition results.(2)Using the above-mentioned tea disease images as the original training samples,new training samples are generated by the unconditional generation model Sin GAN,so as to expand the data.2.A multi-convolution neural network architecture,namely Merge Model,is proposed,and a new weight initialization method is used for automatic identification of tea diseases in small samples.The combination of several CNN modules enables Merge Model to extract various discriminant features,so the recognition ability of Merge Model is improved compared with a single neural network.The convolution filter weights in the network are not randomly initialized,but the image disease features are encoded into them,which helps the model to concentrate on learning useful features at the beginning of training.The experimental results show that Merge Model can effectively distinguish healthy tea from diseased tea,and identify common tea diseases such as tea white star disease,tea leaf blight,tea red spot disease and tea coal disease.Compared with the existing methods,the method proposed in this thesis has higher accuracy of tea disease identification in small samples.3.A tea disease identification algorithm based on cooperative representation learning of image and text in small samples is proposed.Due to the small number of data samples and limited information obtained by using only images,it is difficult to obtain high-accuracy recognition results.Therefore,by adding text information as prior knowledge,the model can obtain good recognition results on small sample data sets.In this thesis,the output content of image information channel and text enhancement channel is combined,and the cooperative ability identification algorithm of disease characteristics is realized by using the correlation degree and complementarity between the two types of information.Image information channel uses visual Transformer to obtain image modal information,and text enhancement channel uses ALBERT combined with Text CNN to obtain text modal information corresponding to tea image.The experimental results show that the new model combining the two modal information can achieve better results than the single image model or text model in small sample data set.更多还原