【作者】 徐丽；

【导师】 姚宇明；

【作者基本信息】 浙江大学 ， 工程硕士（专业学位）， 2021， 硕士

【摘要】 随着社会经济发展水平的提高、人们物质生活水准的提升、人口数量以及城市人口密度的增加,垃圾产出数量也在逐年增加,其对环境的污染、人们健康的影响也日益凸显。针对这个现状,我国开始推行垃圾分类。垃圾分类可充分发挥资源和生产资料的利用价值,减轻人类生产经济活动对自然界的影响。但对于种类繁多的垃圾,如何实现准确的分类,是个棘手的新问题。另外随着人们对深度学习技术的研究不断深入以及硬件算力的不断提升,基于深度学习模型所能解决的问题也越来越多,尤其是在视觉算法模型的应用上,达到了空前的突破。因此,可以基于深度学习技术,设计图像分类模型来解决当前垃圾分类难的问题。本文完成的具体工作如下:首先修改华为垃圾数据集,使其适合本系统的需求。接着设计了以EfficientNet为主干网络,结合注意力模块CBAM的网络模型,CBAM使网络模型将注意力集中在待识别的目标身上,提取出更有效的特征。为解决数据分布不平衡问题以及提升模型的泛化能力,采取了一系列数据增强策略,并修改了模型损失函数,提高模型准确率。然后在分类模型的基础上,设计并实现了C/S架构的微信小程序服务和B/S架构的Web后端管理系统。其中用户可以通过微信小程序提交图像或者检索物体,查询对应的分类结果。针对垃圾识别错误情况,用户可对该图像进行反馈并提交给Web管理端审核。管理员可登陆Web后端系统,审核标注用户反馈的图像以及标注待训练的数据集,用来完成模型的后续优化。最后在验证集和搜集的测试集上对本文设计的分类模型进行实验,实验结果验证了模型的有效性以及较高的准确率;对整个垃圾分类系统进行了功能性和非功能性的测试,结果表明,该系统满足使用需求且具有较高的实用性。更多还原

【Abstract】 With the improvement of the level of social and economic development,the improvement of people’s material living standards,the increase of population and urban population density,the amount of waste output is also increasing year by year,and its impact on environmental pollution and people’s health has become increasingly prominent.In response to this situation,our country has begun to implement waste classification.Garbage classification can give full play to the use value of resources and production materials,and reduce the impact of human production and economic activities on the natural world.But for a wide variety of garbage,how to achieve accurate classification is a thorny new problem.In addition,with the continuous in-depth research on deep learning technology and the continuous improvement of hardware computing power,more and more problems can be solved based on deep learning models,especially in the application of visual algorithm models,which has achieved unprecedented breakthroughs..Therefore,based on deep learning technology,an image classification model can be designed to solve the current difficult problem of garbage classification.The specific work completed in this paper is as follows:First,modify the Huawei garbage data set to make it suitable for the needs of the system.Then designed the EfficientNet as the backbone network,combined with the attention module CBAM network model,CBAM makes the network model focus on the target to be recognized,and extract more effective features.In order to solve the problem of unbalanced data distribution and improve the generalization ability of the model,a series of data enhancement strategies are adopted,and the model loss function is modified to improve the accuracy of the model.Then,based on the classification model,we designed and implemented the We Chat applet service of C/S architecture and the Web back-end management system of B/S architecture.Among them,users can submit images or retrieve objects through the We Chat applet,and query the corresponding classification results.Regarding the error of garbage identification,the user can give feedback on the image and submit it to the web management terminal for review.The administrator can log in to the Web back-end system to review and annotate user feedback images and annotate the data set to be trained to complete the subsequent optimization of the model.Finally,experiments were conducted on the classification model designed in this paper on the verification set and the collected test set.The experimental results verified the effectiveness of the model and high accuracy;the entire garbage classification system was tested functionally and non-functionally.The results show that the system meets the needs of use and has high practicability.更多还原