论文总字数：27670字

摘 要

随着科技的快速发展，人们的生活水平得到了很大的改善，城市生活中的垃圾也在飞速增长。大量的垃圾被运送到城市外填埋或焚烧，仅有部分垃圾进行了无害处理，垃圾的处理速度慢而且垃圾分类智能化水平低。如何实现垃圾的快速分类已成为国家乃至全球不可回避的问题。

本设计主要解决的是现阶段我国居民刚刚接触垃圾分类不久，很多认知习惯还没有养成，对于各种垃圾的类别没有一个清楚的认识，而且在实际的分类场景中不能对垃圾进行准确的分类的现状。通过本地自动分类垃圾桶对家庭及商场场景中少量多次投放的垃圾进行快速分类。分类过程中使用摄像头设备对所需分类垃圾进行视觉识别，通过与MobileNet神经网络生成深度学习模型中的数据集进行匹配确认垃圾类型后对垃圾进行分类投放，同时配合手机APP进行丰富详实的垃圾分类知识的科普。

在实际系统的设计中，本文主要做了以下工作：

1）选取图像采集与处理模块对所需分类的垃圾图像进行获取与输入。选取主控板内嵌深度学习模型与主体程序，根据输入图像的处理数据对垃圾分类模块进行控制。使用垃圾分类模块根据主控板发出信号对所识别垃圾进行对应类型的分类。同时采用容量检测模块和网络连接模块对垃圾桶容量信息进行监测并完成与服务器的通信交互。

2）在对AlexNet、MobileNet、ResNet三种卷积神经网络进行了比较，最终选定了适合移动端嵌入式设备使用的MobileNet。并对MobileNet V1在训练中loss率存在震荡的情况进行了分析和优化处理，提升垃圾识别准确率至可使用范围。

本设计实现的功能如下：

1）使用MaixDuino开发板为主控板，在移动端使用摄像头部件进行图像获取，之后交由图像处理模块与深度学习模型进行识别。按照识别出的垃圾类别通过舵机控制将垃圾倒入四类垃圾类型对应的分类区域内并使用液晶屏显示所识别垃圾的图像信息和识别置信度。使用ESP32芯片通过WIFI与服务器连接，在APP端实现对垃圾桶盖的开关舵机控制。辅以Arduino开发板进行垃圾桶容量数据的采集，通过红外对管和称重传感器配合进行数据获取。并通过Arduino搭载的ESP32芯片实现WIFI连接服务器，将容量数据发送至服务器供客户查询。

2）在使用服务器和安卓APP与硬件系统进行交互时实现了以下三点功能：①可在APP端对垃圾桶进行查找②提供垃圾桶容量数据的展示与实时更新③通过APP对垃圾桶盖进行开关操作。

3）在将垃圾识别置信度设置为85%以上时，对移动端分类库共9类5170个样本进行识别测试，可分为有害垃圾、可回收垃圾、厨余垃圾和其他垃圾4大类，识别准确率达到了97%以上。实现了每类垃圾样本的准确分类。实现了对识别垃圾的类别信息、识别置信度及图像信息的直观展示。实现了本地垃圾桶与网络服务器的连接，能够通过服务器在APP端搜索垃圾桶的名称信息，查看容量信息与控制开关桶盖信号。完成了设计预期的对于智能垃圾桶的数据信息汇总与用户查看垃圾桶可用状态的功能。

本设计最终实现了家庭及商场用少量多次投放垃圾行为的源头垃圾自动分类，帮助居民更加便捷的进行日常生活垃圾的分类，减轻了居民日常分类垃圾的学习和时间成本。同时为垃圾的前端分类实现了自动准确的分类处理，方便了后端垃圾处理企业的回收处理效率，进而提升了垃圾回收行业的经济效益，为我国环保事业中最重要的一环做出贡献。

本设计后续还希望对摄像头进行进一步的升级，以达到更加清晰的图片获取和提升识别准确率的目的。同时对分类算法进行进一步改进，在保证准确率的同时增加模型中包含的类别数量，达到收纳更多种类垃圾数据的目的。

关键词：自动分类垃圾桶；MaixDuino；K210；视觉识别；深度学习

Design and implementation of classified trash can based on machine vision and deep learning

Abstract

With the rapid development of technology, people's living standard has been greatly improved, and the garbage in urban life is also growing rapidly. A large amount of garbage is transported to landfill or incinerated outside the city, and only some of the garbage is treated harmlessly, and the processing speed of garbage is slow and the level of garbage classification intelligence is low. How to achieve rapid waste separation has become an inevitable problem for the country and the world.

This design is mainly to solve the current situation that the residents in China have just contacted with waste classification, many cognitive habits have not yet been developed, and they do not have a clear understanding of the various categories of waste, and they cannot accurately classify the waste in the actual classification scenario. The local automatic garbage cans are used to quickly classify the small amount of garbage that is put out many times in households and shopping malls. In the process of sorting, we use the camera device to visually identify the required sorted garbage, and then confirm the garbage type by matching with the data set in the deep learning model generated by MobileNet neural network, and at the same time, we cooperate with the cell phone APP to provide rich and detailed knowledge of garbage sorting.

In the design of the actual system, this paper mainly does the following work.

1) The image acquisition and processing module is selected to acquire and input the images of the garbage to be classified. The main control board is selected to embed the deep learning model and main program to control the garbage classification module according to the processing data of the input images. The garbage classification module is used to classify the identified garbage according to the signal from the main control board. The capacity detection module and network connection module are also used to monitor the garbage bin capacity information and complete the communication interaction with the server.

(2) After comparing AlexNet, MobileNet and ResNet, MobileNet, which is suitable for mobile embedded devices, is selected, and MobileNet V1 is analyzed and optimized to improve the accuracy of garbage recognition to the usable range.

The functions implemented in this design are as follows.

1) Using MaixDuino development board as the main control board, image acquisition is performed on the mobile side using the camera component, after which it is handed over to the image processing module with the deep learning model for recognition. According to the identified garbage category, the garbage is poured into the classification area corresponding to the four types of garbage through the helm control and the LCD screen displays the image information of the identified garbage and the recognition confidence. The ESP32 chip is connected to the server via WIFI to control the opening and closing of the garbage can lid by the helm in the APP. Supplemented with Arduino development board for garbage can capacity data acquisition, the data acquisition is carried out through the cooperation of infrared pair of tubes and load cells. And through the ESP32 chip equipped with Arduino to achieve WIFI connection to the server, the capacity data is sent to the server for customer query.

2) The following three functions are realized when using the server and Android APP to interact with the hardware system: ① the trash cans can be found in the APP side ② provide the display and real-time update of the trash can capacity data ③ open and close the trash can lid through the APP.

(3) When the confidence level of garbage recognition is set to 85% or more, a total of 5,170 samples of 9 categories are tested in the mobile classification library, which can be divided into 4 categories of hazardous garbage, recyclable garbage, food waste and other garbage, and the recognition accuracy rate reaches 97% or more. Accurate classification of each type of garbage sample was achieved. Visual display of category information, recognition confidence and image information of recognized garbage is realized. The connection between the local garbage can and the network server is realized, and the server can search the name of the garbage can, check the capacity information and control the signal of opening and closing the lid through the server in the APP. The design completes the expected data aggregation of the intelligent trash cans and the user can check the available status of the trash cans.

The design finally realizes the automatic separation of garbage at the source with a small amount of multi-drop garbage behavior in households and shopping malls, helping residents to classify their daily garbage more conveniently and reducing the learning and time costs of daily garbage classification. At the same time for the front-end classification of waste to achieve automatic and accurate classification and processing, to facilitate the back-end waste disposal enterprises recycling processing efficiency, and thus improve the economic efficiency of the waste recycling industry, to contribute to the most important part of China's environmental protection cause.

This design is followed by further upgrades to the camera to achieve clearer picture acquisition and improved recognition accuracy. At the same time, the classification algorithm will be further improved to increase the number of categories included in the model while ensuring the accuracy rate, so as to achieve the purpose of receiving more kinds of garbage data.

Keywords: automatic sorting trash can; MaixDuino; K210; visual recognition; deep learning

目录

摘要 I

Abstract III

第一章 绪论 7

1.1 选题背景 7

1.2 自动垃圾分类的研究意义 7

1.3 国内外研究现状 7

1.3.1国外研究现状 7

1.3.1国内研究现状 8

1.4 本文主要研究内容 9

1.5 本章小结 9

第二章 方案设计与论证 10

2.1 方案概述 10

2.2 本章小结 11

第三章 硬件设计 12

3.1 自动分类垃圾桶的框架设计 12

3.2 基于K210的MaixDuino开发板 12

3.2.1 开发板的组成 13

3.2.2 摄像头模块 13

3.2.3 LCD显示屏模块 14

3.3 舵机模块的原理和功能 14

3.4容量检测模块的原理和功能、 15

3.4.1 红外对管传感器 15

3.4.2 电阻应变式压力传感器与HX711转换模块 16

3.5 本章小结 16

第四章 软件相关理论介绍与设计 17

4.1 传统方法介绍 17

4.1.1 KNN 17

4.1.2 多层感知器 MLP 17

4.2 MobileNet V1 深度学习网络 17

剩余内容已隐藏，请支付后下载全文，论文总字数：27670字