信息工程与人工智能学院知识库

玉米是中国重要的粮食作物，是仅次于水稻的第二大作物，但玉米种植面临 多种病虫害问题的干扰和破坏，主要玉米病害类型包括玉米大斑病、玉米锈病和 玉米褐斑病等，其症状明显表现在玉米叶片上，严重地影响到粮食产业的发展。 目前在利用计算机视觉技术对玉米病虫害进行的研究中，尚无广泛应用的公开、 高质量的玉米病害大型图像数据集，并且目前对于玉米叶片病害的图像识别方法 主要有两种，一是基于图像分割技术提取叶片病斑特征，并结合弱机器学习实现 分类识别，其对数据集数量要求小且训练速度较快，但其分类准确率不高和手工 特征提取比较复杂。二是基于深度学习（常用卷积神经网络）对叶片病斑实现分 类识别，深度学习可以自行提取图像病害特征，识别准确率高且能适应复杂环境 下任务，但有着训练耗时长、模型构建难度较高和技术应用环境要求高等问题。 本文的研究内容如下： （1）建立玉米常见病害图像数据集（Corn Common Diseases Image Dataset， CCDID）。使用 Scrapy 网络爬虫在多个农业技术服务平台爬取玉米种植者上传的 玉米病害图像，结合 PlantVillage 图像数据集中已存在的玉米病害图像在进行人 工审核之后，建立起研究专用数据集 CCDID，包括玉米大斑病、玉米锈病和玉 米褐斑病 3 种玉米常见病害。本文建立的 CCDID 数据集除了自身研究之外，也 可以公开发布以推动玉米病害识别领域的发展。 （2）建立基于深度学习的玉米常见病害图像分类模型 RTA-Net（ResNeSt ThiNet Attention Networks）。本文对 ResNeSt(Split-Attention Networks)卷积神经网 络进行了改进：首先使用迁移学习对 ResNeSt 网络进行预训练；使用 ThiNet 剪 枝神经网络对 ResNeSt 网络进行模型压缩；引入注意力机制选择更好的中间特征 提升分类效果。实验结果表明，RTA-Net 模型相比传统 ResNeSt 模型识别准确率 提高了 2.98%，识别速度提高了 50%。 （3）利用 RTA-Net 模型开发了一款基于微信平台的玉米常见病害图像识别 小程序，具有很高的实践意义，为我国玉米病害识别的智能化研究与应用提供重 要参考价值。微信小程序由用户端、服务端和数据端构成。将训练完的 RTA-Net 模型部署到云平台，在本地环境下设计构建 Flask web 服务器以及数据库系统，兰州财经大学硕士学位论文 基于深度学习的玉米常见病害图像识别研究及应用 用户端选择拍照或者相册上传病害图像至服务端 Flask web 服务器，Flask web 服 务器通过 Restful API 接口访问云平台进行病害识别后返回结果后，通过云平台 返回结果在数据端查找相关病害类型和病害描述，再将病害类型及相关病害描述 返回至用户端进行结果展现，同时服务端将用户上传图像保存至数据端以扩充 CCDID 数据集。经过测试，小程序识别准确率为 90.12%。小程序运行在微信环 境下，操作简单，并能满足用户实时获取病害识别类型的需要。 本文通过建立玉米常见病害图像数据集，采用基于微调（Fine-tune）的迁移 学习方式，然后使用 ThiNet 剪枝网络对 ResNeSt 模型进行压缩,引入注意力机制 搭建新的基于深度学习的玉米常见病害图像分类模型（RTA-Net），经过实验验证 模型取得了较好的识别效果，并开发了基于微信平台的玉米常见病害图像识别小程序。

Corn is an important food crop in China and the second largest crop after rice. However, corn planting is facing the interference and destruction of various diseases and pests. The main types of corn diseases include corn leaf spot, corn rust and corn brown spot, etc. , Its symptoms are clearly manifested on corn leaves, which seriously affects the development of the food industry. In the current research on corn diseases and insect pests using computer vision technology, there is no widely used open, highquality large-scale image data set of corn diseases, and there are currently two main image recognition methods for corn leaf diseases. One is based on images. Segmentation technology extracts leaf lesion features, and combines weak machine learning to achieve classification and recognition. It requires small data sets and fast training speed, but its classification accuracy is not high and manual feature extraction is more complicated. The second is to classify and recognize leaf lesions based on deep learning (commonly used convolutional neural networks). Deep learning can extract image disease features by itself, with high recognition accuracy and adaptability to tasks in complex environments, but it has long training time and difficulty in model building. High requirements and high technical application environment. The research content of this article is as follows: (1) Establish a Corn Common Diseases Image Dataset (CCDID). Use Scrapy web crawler to crawl corn disease images uploaded by corn growers on multiple agricultural technology service platforms, and after manual review with existing corn disease images in the PlantVillage image data set, a research-specific data set CCDID is established, including large corn spots There are three common corn diseases: corn disease, corn rust and corn brown spot. In addition to its own research, the CCDID data set 兰州财经大学硕士学位论文 基于深度学习的玉米常见病害图像识别研究及应用 established in this article can also be publicly released to promote the development of the field of corn disease identification. (2) Establish a deep learning-based image classification model RTANet (ResNeSt ThiNet Attention Networks) for common corn diseases. This article improves the ResNeSt (Split-Attention Networks) convolutional neural network: first use migration learning to pre-train the ResNeSt network; use ThiNet pruning neural network to perform model compression on the ResNeSt network; introduce an attention mechanism to choose a better middle Features enhance the classification effect. The experimental results show that compared with the traditional ResNeSt model, the recognition accuracy of the RTA-Net model is increased by 2.98%, and the recognition speed is increased by 50%. (3) Using the RTA-Net model, a small program for image recognition of common corn diseases based on the WeChat platform was developed, which has high practical significance and provides important reference value for the intelligent research and application of corn disease recognition in my country. The WeChat Mini Program is composed of the user side, the server side and the data side. Deploy the trained RTA-Net model to the cloud platform, design and build the Flask web server and database system in the local environment, the user chooses to take photos or photo albums to upload disease images to the Flask web server on the server, and the Flask web server is accessed through the Restful API interface After the cloud platform performs disease identification and returns the result, the cloud platform returns the result to search for the related disease type and disease description on the data terminal, and then returns the disease type and related disease description to the user terminal to display the result, and the server uploads the image to the user to save To the data end to optimize the model. After testing, the mini program 兰州财经大学硕士学位论文 基于深度学习的玉米常见病害图像识别研究及应用 recognition accuracy rate is 90.12%. The applet runs in the WeChat environment, is simple to operate, and can meet the needs of users to obtain disease identification types in real time. This paper establishes a dataset of common corn diseases image, adopts the transfer learning method based on fine-tune, and then uses the ThiNet pruning network to compress the ResNeSt model, and introduces an attention mechanism to build a new deep learning-based image of common corn diseases Classification model (RTA-Net), the model has been verified by experiments to achieve good recognition results, and a small program for image recognition of common corn diseases based on the WeChat platform has been developed.