【作者】 贺磊；

【导师】 徐慧； 丁允正；

【作者基本信息】 南京林业大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 我国的发展已经进入新时代,新时代下的发展理念是:创新、协调、绿色、开放、共享。绿色发展注重解决人与自然的和谐问题,提倡生活垃圾分类,回收利用垃圾资源,有利于减轻土地侵蚀、环境污染等问题。但由于生活垃圾种类繁多、人们对于垃圾分类标准了解不全面,生活垃圾分类工作往往事倍功半。本文基于改进ResNeXt50网络模型设计一生活垃圾分类系统,使用深度学习的方法用计算机代替人脑对生活垃圾进行分类,帮助人们提高生活垃圾分类工作的效率和质量。本文进行的主要工作如下:(1)针对华为垃圾分类公开数据集中垃圾种类不多、样本图片较少的问题,本文在其基础上制作出一个包含64种垃圾种类、30848张样本图片的生活垃圾分类数据集;并搭建出AlexNet、VGG16、GoogLeNet、ResNet50这四种主流的图像分类网络模型,在生活垃圾数据集上对比训练效果,选定分类效果更好的ResNet系列,在其基础上搭建出ResNeXt模型,该模型的分类准确率可达80%;(2)针对ResNeXt基础模型分类准确率不高的问题,本文做出如下改进:在ResNeXt模型50-layer和101-layer两种结构中使用基于ImageNet数据集的迁移学习技术,网络训练结果表明迁移学习可使ResNeXt50在生活垃圾数据集上达到更高的分类准确率;为增强模型提取图像特征信息的能力,将SENet中的注意力机制融入ResNeXt50的残差网络结构中,搭建出SE-ResNeXt50模型,进一步提高了模型的分类准确率;为加快网络模型训练和收敛的速度,提高模型的训练效果,本文还提出了一种具有指数型动态学习率的Adam优化算法,并在模型的训练中验证了效果,最终设计出的SE-ResNeXt50-lr网络模型在生活垃圾数据集上的分类准确率达到96.23%,可以对垃圾图像进行精准的识别与分类;(3)针对使用者无法直接操作网络模型进行垃圾分类的问题,本文使用Tkinter库设计了垃圾分类系统的可视化前端操作界面,对后端网络模型进行操控和数据传输,并对系统前后端分别进行了功能调试。本文最终设计出基于SE-ResNeXt50-lr网络模型的生活垃圾分类系统,并通过现实环境的应用测试验证了系统的可行性与实用性,该系统能够帮助人们更好的解决生活垃圾分类问题,达到课题预期目标,满足实际应用需求。更多还原

【Abstract】 China’s development has entered a new era,in which the concept of development is innovation,coordination,green development,openness and sharing.Green development focuses on solving the problem of harmony between man and nature,advocating the classification of household waste and recycling garbage resources,which is conducive to reducing land erosion and environmental pollution.However,as there are many kinds of household waste and people do not have a comprehensive understanding of the household waste classification standards,household waste classification often gets half the result with twice the effort.This paper designs a household waste classification system based on the improved ResNeXt50 network model,and uses deep learning method to classify garbage by computer instead of human brain,so as to help people improve the efficiency and quality of garbage classification.The main works of this paper are as follows:(1)In view of the problem that there are not many garbage types and few sample pictures in huawei garbage classification public data set,this paper makes a household garbage classification data set with 30848 sample pictures of 64 garbage types on its basis.Four mainstream image classification network models,AlexNet,VGG16,GoogLeNet and ResNet50,were built.The training effect was compared on the household waste classification data set,and the ResNet series with better classification effect was selected.On this basis,the ResNeXt model is built,and the classification accuracy of the model can reach 80%;(2)In view of the low classification accuracy of ResNeXt basic model,this paper makes the following improvements: In ResNeXt model 50-Layer and 101-Layer,transfer learning technology based on ImageNet data set is used.The network training results show that transfer learning can make ResNeXt50 achieve higher classification accuracy on the household waste classification data set.In order to enhance the ability of extracting image feature information from the model,the attention mechanism in SENet was integrated into the residual network structure of ResNeXt50 to build the SE-ResNeXt50 model,which further improved the classification accuracy of the model.In order to accelerate the training and convergence speed of network model and improve the training effect of the model,this paper also proposed a Adam optimization algorithm with exponential dynamic learning rate,and verified the effect in the training of the model.Finally,the classification accuracy of SE-ResNeXt50-lr network model designed on the household waste classification data set reached 96.23%.Garbage image can be accurately identified and classified;(3)To solve the problem that users cannot directly operate the network model for garbage classification,this paper uses Tkinter library to design the visual front-end operation interface of the household waste classification system,control the back-end network model and transmit data,and debug the front and back ends of the system respectively.This paper finally designs the household waste classification system based on the SE-ResNeXt50-lr network model,and verifies the feasibility and practicability of the system through the application test in the real environment.The system can help people better solve the household waste classification problem,achieve the expected goal of the project,and meet the actual application requirements.更多还原