A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE Preprint compiled October 4 2022

2025-04-27 0 0 1.82MB 12 页 10玖币
侵权投诉
A SMART RECYCLING BIN USING WASTE IMAGE
CLASSIFICATION AT THE EDGE
Preprint,compiled October 4, 2022
Xueying Li and Ryan Grammenos
candy.li.19@ucl.ac.uk , r.grammenos@ucl.ac.uk
Electrical and Electronic Engineering
University College London
Abstract
Rapid economic growth gives rise to the urgent demand for a more efficient waste recycling system.
This work thereby developed an innovative recycling bin that automatically separates urban waste
to increase the recycling rate. We collected 1800 recycling waste images and combined them with
an existing public dataset to train classification models for two embedded systems, Jetson Nano and
K210, targeting different markets. The model reached an accuracy of 95.98% on Jetson Nano and
96.64% on K210. A bin program was designed to collect feedback from users. On Jetson Nano, the
overall power consumption of the application was reduced by 30% from the previous work to 4.7
W, while the second system, K210, only needed 0.89 W of power to operate. In summary, our work
demonstrated a fully functional prototype of an energy-saving, high-accuracy smart recycling bin,
which can be commercialized in the future to improve urban waste recycling.
I. Introduction
Waste generation has increased dramatically in the 21st century
due to the growth in the global population. According to
the World Bank Group’s report [
1
], 2.01 billion tonnes of
municipal solid waste are generated worldwide every year,
with only about 19% adequately recycled [
1
]. Recycling not
only helps to preserve raw material but, more importantly,
reduces the landfills required, which is an undesirable way
of waste disposal due to its high demand for space and the
danger of introducing contaminants into the ground or even
groundwater system. A major step of recycling is to separate
the waste into specific categories according to its material.
Failure to do so will significantly harm the effectiveness of
recycling. Traditionally, workers at the recycling company sort
the waste into corresponding categories by hand, which is an
inefficient method and requires unnecessary labor. Therefore,
people are now in dire need of a more advanced and automated
waste separation system.
Image classification machine learning (ML) algorithms have
been used to build automatic waste classification systems and
assist waste management. This research aimed to improve the
existing classification system and pursue the commercialization
of an artificial intelligence (AI) street bin. Specifically, we
focused on reducing the power consumption of the controller
board to extend AI bin’s battery life, increasing its waste
classification accuracy and reducing its price.
The only intelligent recycling bin on the market, Bin-e [
2
],
uses image classification algorithms to separate the trash into
four categories: plastic, paper, metal and glass. It achieves
a waste segmentation accuracy of 92% [
2
]. However, the
price USD 5200 [
2
] is expensive for a rubbish bin. Also,
Bin-e’s operation requires a 230V power supply, so it cannot
replace the traditional trash bins at the locations without plug
sockets. Therefore, it can hardly increase the household and
street recycling rate. As a result, this paper proposes low-cost
battery-powered real-time recycling waste segmentation bin
systems to fill this gap.
In this research, we trained a light MobileNet image classifica-
tion model for Jetson Nano with TrashNet [
3
], an open-source
waste dataset, and 1800 new training samples collected by
us. The model demonstrated great performance with a high
test accuracy of 95.98% and a small parameter size of 3.0 M.
Jetson Nano only consumed 4.7 W when it ran the model. We
also found a cheaper and less energy-hungry device, K210
[
4
] to further reduce power consumption. The K210 board
only consumed 0.89 W at inference time and the model on
it achieved a high test accuracy of 96.64%, which made it a
more practical solution for trash bin applications. The code
used in this study is available on Github [5].
This paper is organized as follows. In Section II, the related
works performed in waste classification and AI trash bins
are introduced. Section III introduces the system design and
explains the background theories that support this study.
The methodology of data collection and model training is
introduced in section IV. Section V analyses the test results
and evaluates the system performance on Jetson Nano and
K210. Finally, Section VI concludes the achievements and
provides suggestions for future research.
II. Related works
The United Kingdom (UK) government has planned to increase
the household recycling rate in England to 50% by 2020,
but only 44% of municipal waste was reused and recycled
in 2020 [
6
]. The inconvenience and lack of knowledge are
arXiv:2210.00448v1 [cs.CV] 2 Oct 2022
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 2
two crucial factors that prevent people from recycling [
7
].
Automatic waste segmentation bins were designed to overcome
these problems by helping people classify and send the waste
into the corresponding containers, making waste disposal more
convenient. Table I outlines the approaches and the primary
hardware components involved in related studies to construct
the waste segmentation systems.
In early research, a microcontroller is connected to a variety of
sensors to determine the composition of the waste. For example,
inductive and capacitive sensors can detect the metal element
[
8
][
9
][
10
] and moisture sensors separate wet waste from dry
waste [
9
][
10
]. The microcontroller makes decisions based
on its readings [
8
][
9
]. However, although the sensor-based
classification method can detect the composition precisely, it
lacks the ability to classify waste into more specific groups. The
sensors cannot discriminate between plastic, paper, glass, and
other unrecyclable dry waste, which are important categories
in recycling.
The development of machine learning and image classification
enables the bin to sort the waste based on visual input
like a human. The convolutional neural network (CNN) is
a branch of image classification algorithms that performs
mainly convolution operations on the pixels [
15
]. It is the most
popular choice due to its high accuracy and power efficiency
compared to other methods and is used in all the four papers
[
11
][
12
][
13
][
14
]. It gives the bin ability to differentiate between
spoons and cups, which is tremendous progress compared to the
sensor-based approaches. Traditionally, the CNN models run on
a cloud raising the data transmission latency and user privacy
security problems. To solve these problems, recent research
moved the computation to an edge embedded system. However,
edge computing has the drawback of limiting computation
resources, so the model size is important in selecting the CNN
model structure.
To develop, evaluate and select the proper CNN structures, most
research in this field used the TrashNet dataset developed by
Mindy Yang and Gary Thung [
3
] in 2017, the first open-access
recycling waste dataset. This high-quality dataset contains
2527 waste photos from 6 groups: metal, glass, plastic, paper,
cardboard, and trash. It provides a foundation for later research
and our study also used this dataset.
Efforts have been made to increase the segmentation accuracy
of CNN models based on TrashNet. The state-of-art accuracy
of 97.86% on TrashNet was reached in [
16
]. However, the
CNN architecture, GoogLeNet, used in this research will cause
out-of-memory (OOM) on Jetson Nano [
17
]. It demonstrated
the potential of CNN classification, but the model size and
computation cost must be cut down to implement the machine
learning algorithm on embedded systems. A lighter CNN
model, WasteNet, was constructed by White et al. and achieved
an accuracy of 97% on the TrashNet dataset [
12
]. The paper
claimed that the model could be loaded to Jetson Nano, but
it did not provide details regarding the edge implementation,
such as the classification speed. Nevertheless, from the model
structure, we could estimate that the classification speed for
one image would be too slow for real-time classification. It
can only be used in a bin application that classifies the waste
objects based on one photo.
The classification speed is quantified by inference time, which
refers to the time taken for one classification to be completed.
Our application looks for a smaller model that can interact
with the user in real-time, with an inference time that must be
smaller than 0.1 s, which is the human visual reaction time. The
CNN model, EfficientNet B0, used in [
17
] is the starting point
of this research and it achieved an accuracy of 95.38% and
an inference time of 0.07 s on Jetson Nano. While the model
is fast enough to be used in real-time applications, the 96%
high memory usage needs optimization for bin applications.
Power consumption is another factor that needs to be consid-
ered in the final product. Unfortunately, rare research has paid
attention to it. For instance, none of the trash bin cases listed in
Table I has measured the power consumption of the proposed
system. Still, it is evident that the CNN-based approaches have
higher power consumption than the sensor-based one and the
Pynq-Zl and Raspberry Pi will typically have lower power
consumption than Jetson Nano used in [
12
] and [
17
]. The
Raspberry Pi 4 has a typical power consumption of between
2.7 W and 6.4 W. It reduces the power consumption but
results in undesirable low performance [
14
]. As a result, the
CNN architecture, MobileNet V2, has a low average per-class
precision of 91.76% and a long inference time of 0.358 s on
it.
The limitation of the current research is that the energy-
saving systems will have unacceptable low performance in
the classification task for commercialization. As a result, this
study designed two high-accuracy waste classification systems
with lower power consumption than previous studies. The first
system reduced the power consumption of the application on
Jetson Nano by using a lighter model, MobileNet, to reduce
power consumption while maintaining the accuracy.
The second system is developed based on K210, a less
expensive and more energy-saving embedded device. K210 is
an unpopular choice and is used in fewer than 100 research.
Most research implemented YOLO object detection models on
it [
18
][
19
] and demonstrated its outstanding power efficiency.
K210 has not been used for recycling waste classification when
this paper was written, so this paper proposed and evaluated
an innovative approach.
III. System design
A. The AI recycling bin design
The AI bin consists of five recycling waste containers and a
detection box. The waste will be sorted after it is placed in
the detection box. The whole system can be controlled by a
Jetson Nano or a K210 board.
The bin design using Jetson Nano is summarized in Figure
1. Jetson Nano will interact with the users and collects
feedback through the touch screen in front of it, displaying
the instructions for using the bin and the camera inputs. The
Raspberry Pi camera at the top takes photos of the waste in
the detection box. The images will be fed to the classification
algorithm in Jetson Nano, which classifies the photos into
seven groups. The waste classification models in the previous
study [
17
] have five output classes: "paper", "metal", "plastic",
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 3
TABLE I
Related works in AI bins
Year Author Waste classification method Classification category The device
2014
Rajkamal et al.
[8]
Inductive metal sensor,
capacitive based moisture sen-
sor, methane sensor, odor sen-
sor
metal/glass, food, bio,
paper/plastic, inert
PIC18F4580
controller
2017
kumar et al. [
9
]
Inductive and capacitive metal
sensor, gas sensor, bacteria sen-
sor, moisture sensor
Biodegradable, plastic, metal STM32 controller
2019
Pereira et al.
[10]
Infrared radiation sensor, capac-
itive sensor
Dry, wet, plastic Atmega 328P
microcontroller
2019
Ziouzios et al.
[11]
CNN image classfication
Glass, paper, metal, plastic,
cardboard, trash
Xilinx Pynq-Zl FPGA
2020
White et al.
[12]
CNN image classification Paper, plastic, metal, glass,
cardboard, other
Jetson Nano
2021
Jimeno et al.
[13]
CNN image classification
Aluminum cans, plastic
bottles, plastic cups, paper,
spoons/forks
Computer with
Nvidia RTX 2060 GPU
2021
Sallang et al.
[14]
CNN image classification
Glass, paper, metal, plastic,
cardboard
Raspberry Pi 4
Fig. 1. System block design of WasteNet.
"cardboard" and "glass". They will be reproduced in the next
section as the benchmark models of our study. Our model will
be trained with two new classes, "empty" and "hand". "Empty"
means there is nothing in the photo, while the "hand" group
detects a human’s hand to avoid trapping the user’s hand by
the door. When these two groups are detected, the bin waits
and continues detecting.
B. The Benchmark models
The waste classification models in [
17
] consist of an input
layer, a pre-processing augmentation layer, an EfficientNet B0
base model layer, a global average pooling 2D layer and a
dense layer. The first model has an input layer size of 512x384
pixels, while the second model input has 384x288 pixels. Table
II shows the configuration of the augmentation layer and the
model training hyperparameters for reproduction.
The models were trained with the TrashNet dataset, which
consists of six categories of images: glass, paper, cardboard,
plastic, metal, and trash. Only the first five groups of data
were used to build the model that classified recycling into
five categories. In total, 2152 images were used for training
TABLE II
Training parameters used in [17]
Pre-Processing Method Value
Random flip Horizontal and vertical
Random rotation Up to 180◦
Random translation Up to 10%
Random zoom Up to 75%
Training Parameter Value
Learning rate scheduler Constant learning rate scheduler
Train/Validation/Test split ratio 72/18/10
Optimizer Adam optimiser
Training epochs 50
Learning rate 4.3e-05
Fine-tuning training epochs 8
Fine-tuning learning rate 4e-06
Loss functions Sparse categorical cross entropy
Classifier activation function Softmax
Include top layers in base model False
Batch size 16
and validation, and 238 images were used for testing. The
train/validation/test split is 72/18/10.
To begin with, the base model is initialized with pre-trained
weights on ImageNet. The two models are trained separately
by setting the corresponding input sizes. The learning rate
was set to 4.3e-05 at the first 50 epochs and was reduced to
4e-06 at the eight epochs afterwards. Eventually, both models
achieved the same test accuracy of 95.38%. Figure 2 shows
the confusion matrix of the test results and indicates that most
mistakes are made in classifying between paper and cardboard.
摘要：

ASMARTRECYCLINGBINUSINGWASTEIMAGECLASSIFICATIONATTHEEDGEPreprint,compiledOctober4,2022XueyingLiandRyanGrammenoscandy.li.19@ucl.ac.uk,r.grammenos@ucl.ac.ukElectricalandElectronicEngineeringUniversityCollegeLondonAbstractRapideconomicgrowthgivesrisetotheurgentdemandforamoreecientwasterecyclingsystem....

展开>> 收起<<
A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE Preprint compiled October 4 2022.pdf

共12页,预览3页

还剩页未读， 继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

相关推荐

更多
立即下载
分类：图书资源 价格：10玖币 属性：12 页 大小：1.82MB 格式：PDF 时间：2025-04-27

开通VIP享超值会员特权

  • 多端同步记录
  • 高速下载文档
  • 免费文档工具
  • 分享文档赚钱
  • 每日登录抽奖
  • 优质衍生服务

作者详情

相关内容

更多

热门标签

人际关系 动力学连接体 力的合成 配电装置 高考理综 全宋诗作者索引 公务员考试
/ 12
评分 收藏
分享
立即下载
客服
关注