Global Journal of Computer Science and Technology: G

Interdisciplinary
Volume 19 Issue 4 Version 1.0 Year 2019
Type: Double Blind Peer Reviewed International Research Journal
Publisher: Global Journals
Online ISSN: 0975-4172 & Print ISSN: 0975-4350

Using Deep Learning to Detect Polyethylene Terephthalate

(PET) Bottle Status for Recycling
By Hellen Wanini Mwangi & Mpai Mokoena
Abstract- Following the recent ban of plastic waste import by China, many developed countries
face challenges with their huge amount of plastic waste. Some of the countries have diverted
their waste to other developing East-Asia countries like Philippines, Vietnam and Malaysia.
However, Malaysian government has taken strict action to send back over 3000 tons of plastic
waste to its origin citing contamination. The aim of this paper is to establish mechanisms to
detect the status of post-consumer PET bottles for recycling. By encouraging recycling of clean
pet bottles, we ensure high quality bottles for recycling. A research based as well as
experimental design approach was adopted to develop mechanisms to detect PET bottle status.
During the experiment, various PET bottles were collected and images captured. A total of 1749
images were captured using raspberry Pi camera.
Keywords: recyclable pet bottles, squeeze net, recycling, predictions, deep learning, incentives.
GJCST-G Classification: I.2.6

UsingDeepLearningtoDetectPolyethyleneTerephthalatePETBottleStatusforRecycling

Strictly as per the compliance and regulations of:

© 2019. Hellen Wanini Mwangi & Mpai Mokoena. This is a research/review paper, distributed under the terms of the Creative
Commons Attribution-Noncommercial 3.0 Unported License http://creativecommons.org/licenses/by-nc/3.0/), permitting all non
commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
 Using Deep Learning to Detect Polyethylene
Terephthalate (PET) Bottle Status for Recycling
Hellen Wanini Mwangi α & Mpai Mokoena σ

Abstract- Following the recent ban of plastic waste import by by many developing and developed countries to recycle
China, many developed countries face challenges with their the plastic PET bottle waste.

2 019
huge amount of plastic waste. Some of the countries have According to The Guardian 2017, annual
diverted their waste to other developing East-Asia countries

Year
production of plastic bottles is expected to increase
like Philippines, Vietnam and Malaysia. However, Malaysian
from 500 billion to half a trillion by 2021(figure1).
government has taken strict action to send back over 3000
tons of plastic waste to its origin citing contamination. The aim 27
of this paper is to establish mechanisms to detect the status of

Global Journal of Computer Science and Technology ( G ) Volume XIX Issue IV Version I
post-consumer PET bottles for recycling. By encouraging
recycling of clean pet bottles, we ensure high quality bottles
for recycling. A research based as well as experimental design
approach was adopted to develop mechanisms to detect PET
bottle status. During the experiment, various PET bottles were
collected and images captured. A total of 1749 images were
captured using raspberry Pi camera. The images belong to 4
different classes; seal, cap, seal & cap, no seal, cap &
content. Deep Learning technology (Squeeze Net) was used
to train the PET bottle images. After training the bottle images,
the model achieved 98% accuracy with correct bottle status
recognized. The trained model is then deployed on a Figure 1: PET bottle production
raspberry pi to detect PET bottles in real time when using a
reverse vending machine. When measuring prediction a) Problem statement
performance, the model showed a delay of 0.018 to 0.022 PET is the most widely produced, consumed
seconds per prediction using Intel CPU. Whereas measuring and recycled plastic in the world. Although PET is
the prediction performance in raspberry pi which is 5 to 10 recyclable, a huge quantity is discarded. Consequently,
times slower than the Intel CPU with a delay of 0.1 to 0.25 the huge amount of PET that is not recycled poses a
seconds per prediction.
threat to the environment, negatively affecting marine life
Keywords: recyclable pet bottles, squeeze net, recycling,
and human health. As a result, this paper proposes a
predictions, deep learning, incentives.
solution that can ensure the quality PET bottles are
I. Introduction collected and recycled.

C
urrently, recyclers are looking for more from b) Objectives
shredders. According to Plastic Recycling Based on the topic, the paper is guided by
Technology 2019, new ways to boost productivity, several research objectives:
quality and profitability are being sought. There is more • To establish effective mechanisms for post-
demand for quality Polyethylene Terephthalate (PET) consumer PET bottles status recognition in real-time
bottles. “Purity rates of 98% are now being sought as using reverse vending machines.
standard, which means there’s no longer a place for • Build prediction system to recognize PET bottle
contaminated plastics PET bottles. Furthermore over a status using sensors and deep learning.
million plastic bottles are bought every minute globally, • To test and evaluate the performance of the deep
by 2021 this number is expected to increase by 20%. learning model in bottle status recognition
This means firms have to prioritize quality PET bottles
through reduce, reuse, recycle principles.” c) Methodology
Consequently, when the three principles of recycling are A qualitative and quantitative method as well as
not followed, over 8 million tonnes of plastic bottles experimental method was used. The reason for
enter the sea every year (Plastic Oceans, 2017). Thus, selecting these methods. Both qualitative and
efforts to collect the plastic waste have been expressed quantitative methods were used to analyze the problem.
Experiment was conducted by collecting post-consumer
Author α σ : e- mails: [email protected], [email protected] PET bottles to train the model. The experiment helped to

© 2019 Global Journals

 Using Deep Learning to detect Polyethylene Terephthalate (PET) Bottle status for Recycling

verify the solution which is the identification of pet bottle According to the interview results, the
status using deep learning. consumers return the bottles to the drop-off points
This paper is structured as follows: Section 2 where the vending machines are located. However, the
covers recycling frameworks; Section 3 compares system does not take into consideration the quality of
existing solutions; Section 4 discusses the proposed bottles returned such as the removal of cap, label or
solution; Section 5 is about the prototype development; content or cleanliness. Upon returning the bottles,
Lastly, in Section 6 the results are discussed. consumer gets a coupon which can be redeemed at the
stores. The vending machine scans the barcode of the
II. Recycling Frameworks bottle and it rejects any bottle that is not Swedish. As a
a) SDGs result, the system encourage recycling of PET bottles by
providing rewards.
2 019

The term SDGs refer to the Sustainable

From the research findings, Sweden deposit-
Development Goals set by the United Nations
refund system is mandatory, it began in the 1970s in
Year

Development Program (UNDP).Several of these touches

order to achieve the 75% recycling rate indicated by the
on waste management, health and marine life.
38
28 government. The key stakeholders of the system include
the can manufacturer, breweries and retailers who
established a company by the name Returpack to set
Global Journal of Computer Science and Technology ( G ) Volume XIX Issue IV Version I

up collection and recycling system for aluminium cans.

Also, Returpack-PET (AB Svenska Returpack PET) was
established for PET bottles in 1994. (Tojo, 2011). Prior to
Sweden having its own PET recycling plant, the PET
bottles were sent to Germany for cleaning and recycling.
Source: UNDP In an effort to reduce PET bottle waste, a recycling
Figure 2: Sustainable Development Goals this paper company was established in 2006-2007. As a result,
touches on Sweden achieved a recycling rate of 90% for pet bottles
By collecting post-consumer PET bottles for (The Locals.Se, 2018)
recycling, we ensure a clean environment, marine life as c) PETCO South Africa
well as good health from a clean environment. It also PETCO is a South African company which was
ensures that there is responsible production and incorporated in 2004 to self-regulate recycling of post-
consumption of PET bottles produced and end-of-life. consumer PET bottles. PETCO is guided by the principle
The principle of Extended Producer Responsibility of Extended Producer Responsibility. It Involves various
guides manufacturers to ensure responsible production. stakeholders in its business model which makes it a
III. Existing Solutions success. Stakeholders involved are raw material
producers, converters, brand owners', retailers,
a) Overview consumers, recyclers and donors. Extended producer
In this section, the paper compares several responsibility ensures the end of life of the products
case studies in relation to plastic PET bottle waste which the companies produce.
management in different countries. The cases are based The country’s recycling rate exceeded
on: 1) Successful models of reverse vending machines expectations in increasing from 52% in 2015 to 55% in
for post-consumer PET bottle collection, case of 2016(Post & Number, 2018). In 2017, 2 billion PET
Sweden; 2) The PETCO model by Plastic Recycling bottles were recycled, the recycling rate was 65%.
Company of South Africa; 3) The case of Japan. PETCO South Africa has embraced a circular economy
b) Returpack Sweden paying much attention to achieving the Sustainable
The Returpack case study on Deposit Refund Development Goals (SD Goals: 14, 12, 6, and 3) as
Systems is insightful. The system takes care of four shown in figure 2.
types of beverage containers PET bottles being one of Another contribution to PETCO South Africa
them. The collection systems promote the collection of success in recycling is the existence of state-of-the-art
recyclables and waste with the provision of economic recycling facilities. Also, recyclers can buy PET bottles
incentives. There are established deposit refund from collectors with the fee collected from the voluntary
systems in Sweden and reward points collection fee paid by importers and manufacturers of PET bottle
systems in place. To have a clear understanding of how products.
the deposit refund systems work in Sweden, a research In a 2017 report by Plastics SA, the major
on Sweden Deposit Refund System has been done by challenge faced in the recycling of plastic bottle waste is
way of interview and obtained first-hand information getting access to good quality, clean bottles before they
from a resident citizen of Sweden. reach the landfills.

© 2019 Global Journals

 Using Deep Learning to detect Polyethylene Terephthalate (PET) Bottle status for Recycling

d) Plastic Recycling in Japan Table 1: Originality

Waste management in Japan is successful due
to waste separation practice. For example, the first
house rules for a tenant from a landlord is about waste
separation. Kamikatsu in Japan has a collection centre
for recyclables. The consumers return bottles and are
awarded points. After the points accumulate, the
residents can enjoy discounts for basic services such as
housing, water, health etc.

IV. Proposed Solution

2 019
In this paper a reverse vending machine with

Year
capability to identify bottle status real-time is proposed. V. Solution Development
This is the contribution to already existing solutions in
collecting post-consumer PET bottles. The purpose of a) System configuration 29
the solution is to encourage the return of recyclable In the reverse vending machine, we have a

Global Journal of Computer Science and Technology ( G ) Volume XIX Issue IV Version I
quality PET bottle which range/object sensor, a camera sensor and controller
raspberry pi. When the bottle is put in the machine, the
a) How it works
range sensor detects the bottle and triggers the camera
When the consumers finish with the pet bottles,
to take photos. Figure 4 shows an overview of reverse
they clean and return the empty bottles to a collection
vending machine with PET bottle status recognition.
centre/drop-off point, there the bottle is put in the
reverse vending machine, the bottle status is identified
based on the following features; With seal or without
seal, with cap or without cap, with cap and seal, or if it
has no cap, content and seal. Awarding of incentives
depends on the status and for each status, the amount
awarded varies.
Figure 3 shows bottle status recognition which
varies from low incentives for not clean bottles to high
incentives for clean ones.

Figure 4: Reverse Vending Machine

In the taking photo process of training images,
the images are saved in a network server (Samba)
Figure 3: Bottle Status Recognition directory created on Linux computer.
Using deep learning technology to detect the During the prediction process, the bottle put in
status of the PET bottle in real-time in a reverse vending the reverse vending machine is captured and its image
machine and give different amount of incentive as is saved on SD card in the raspberry pi. Using deep
shown on the figure above. learning model (Squeeze Net) deployed on the
raspberry pi, the bottle image is predicted and results
b) Comparison and Originality of the proposed solution displayed on the dashboard as shown in figure 5 below.
to existing solutions
Table 1 shows a comparison of the proposed
solution to existing solutions. The originality of the
proposed solution is 1) it adopts divergent incentives
(the amount of incentives is decided based on the bottle
status), and 2) the bottle status is recognized
automatically.

© 2019 Global Journals

 Using Deep Learning to detect Polyethylene Terephthalate (PET) Bottle status for Recycling

Figure 6 shows the file directory structure of the

image data for deep learning training. The directories
“1000”, “1100”, “0100” and “0000” correspond to the
bottle status a) to d). The captured image data is
automatically saved to the corresponding directory. The
images are saved in respective directories according to
their classes. For example images of seal, cap, cap and
seal are saved in 1000, 0100, and 1100 respectively.
2 019Year

38
30
Global Journal of Computer Science and Technology ( G ) Volume XIX Issue IV Version I

Figure 5: System Configuration Diagram

b) Hardware configuration Figure 6: Dataset organizational structure
The configuration involves several devices.
There are requirements in order to configure all the VI. Results and Discussions
devices to work well. Below is a list of the requirements: a) Performance of Deep Neural Network
1. Range sensor (HC-SR04) Squeeze Net (Iandola, Forrest N., et al, 2016) is
2. Raspberry pi camera selected for the neural network because it is small
3. The raspberry pi 2 model B enough to be executed on Raspberry pi. Using the
4. PCB (A circuit board) Squeeze Net model successfully trained the dataset.
5. Wires Total parameters are 541,332. Training parameters are
6. Four LEDs 540,812 and non-trainable parameters are 540. Some of
7. Four toggle switches corresponding to each LED the hyper parameters are learning rate, weight decay
(L2), 20 epochs and the ‘accuracy’ metrics to show the
Four toggle switches are connected to four results. Figure 7 shows the loss and accuracy during the
GPIOs of Raspberry pi, so that the software on training going on
Raspberry pi can sense the status of the switches. This
is used in the function of taking images for the training
datasets. Four LEDs are connected to other four GPIOs
of Raspberry pi, so that the software on Raspberry pi
can turn on or off individually. This is used in the function
of taking images for prediction to display the prediction
result.
c) Generating Training dataset
The training dataset requires a label which
corresponds to the status of the bottle for each image of
the bottle, because the supervised machine training is
used. In the prototype, two toggle switches out of four
are used to indicate the label of the image. The status of
bottles which must be recognized are the existence of
seal, cap. So the number of bottle status is four: a) only
seal exists, b) both seal and cap exist, c) only cap
exists, and d) no seal and no cap exists. The four
statuses are coded as binary values as a) 1000, b)
1100, c) 0100, d) 0000. Each binary digit is assigned to
each toggle switch, so the user of the prototype easily
specify the state of the bottle by setting the toggle
switches. This feature is useful for speeding up the
training dataset generation process.
Figure 7: Visualizing training process
© 2019 Global Journals
 Using Deep Learning to detect Polyethylene Terephthalate (PET) Bottle status for Recycling

Loss
Networks with transfer learning for computer vision-
Model loss is a value that tells us how well or
based data-driven pavement distress detection.
bad our model classifier performed at each iteration.
Construction and Building Materials, 157, pp.322-
The lower the loss value the better the model performed
330.
and vice versa. It is easy for people to predict images
6. Iandola, F., & Keutzer, K., 2017, October. Small
with their human eyes but in deep learning it is difficult
neural nets are beautiful: enabling embedded
due to various challenges. Therefore, the need to apply
systems with small deep-neural-network
a loss function in our model. There is no universal loss
architectures. In Proceedings of the Twelfth
function that works for all classifications.
IEEE/ACM/IFIP International Conference on
Accuracy Hardware/Software Codesign and System Synthesis
Accuracy of a model is determined after Companion (p. 1). ACM.

2 019
learning has taken place. Then the test samples are fed 7. Iandola, F. N., Han, S., Moskewicz, M. W., Ashraf,
to our model to determine the accuracy. K., Dally, W. J. and Keutzer, K., 2016. Squeeze Net:

Year
Accuracy = No of correct predictions / Total no Alex Net-level accuracy with 50x fewer parameters
of predictions and< 0.5 MB model size. Ar Xiv preprint 31
The accuracy in the training of pet bottles is arXiv:1602.07360.
83.7 % for 20 epochs .After increasing the number of 8. Jambeck, J. R., Geyer, R., Wilcox, C., Siegler, T. R.,

Global Journal of Computer Science and Technology ( G ) Volume XIX Issue IV Version I
epochs, training accuracy improved. The results are Perryman, M., Andrady, A., Narayan, R. and Law,
98% accuracy. K.L., 2015. Plastic waste inputs from land into the
ocean. Science, 347(6223), pp.768-771.
b) Measuring the prediction performance
9. LeCun, Y., Bengio, Y. and Hinton, G., 2015. Deep
The prediction results shown in table 2 shows
learning. nature, 521(7553), p.436.
the prediction performance on Intel CPU which were
10. Pi, W., 2017. GPIO Interface library for the
0.022 seconds in the 1st test and 0.018 seconds in the
Raspberry Pi. URL: http://wiringpi. com/(visited on
2nd test. Since the raspberry pi CPU is 5 to 10 times
2017-07-13).
slower than the Intel CPU, the approximated delay time
11. Rawat, W. and Wang, Z., 2017. Deep convolutional
is 0.1 seconds to 0.25 seconds per prediction.
neural networks for image classification: A
Table 2: Squeeze Net performance results in predicting comprehensive review Neural computation.
the images 12. Rawat, W. and Wang, Z., 2017. Deep convolutional
neural networks for image classification: A
comprehensive review. Neural computation, 29(9),
pp.2352-2449.
13. Setting up a samba server as a stand-alone server.
Retrievedfrom:https://wiki.samba.org/index.php/Sett
ing_up_Samba_as_a_Standalone_Server
14. https://www.theguardian.com/environment/2017/jun
/28/a-million-a-minute-worlds-plastic-bottle-binge-
as-dangerous-as-climate-change
15. Voulodimos, A., Doulamis, N., Doulamis, A. and
Protopapadakis, E., 2018. Deep learning for
computer vision: A brief review. Computational
References Références Referencias intelligence and neuroscience, 2018.
16. Arun David Ambrose., 2017. PET recycling - a
1. Bengio, Y., 2009. Learning deep architectures for AI. means to save the planet.
Foundations and trends® in Machine Learning, https://www.beroeinc.com/whitepaper/pet-recycle/
2(1), pp.1-127. 17. The Locals se., 2019. That's pant! The story behind
2. Camera Exposure matrix. Retrieved Sweden's bottle recycling scheme. Retrieved from:
from:https://publiclab.org/notes/MaggPi/07-02- https://www.thelocal.se/20180328/thats-pant-the-
2018/camera-exposure-matrix story-behind-swedens-bottle-recycling-system
3. Chollet, F., 2015. Keras: Deep learning library for
theano and tensor flow. URL: https://keras. io/k,
7(8), p.T1.
4. Gay, W., 2018. Pi Camera. In Advanced Raspberry
Pi (pp. 493-499). Apress, Berkeley, CA.
5. Gopalakrishnan, K., Khaitan, S. K., Choudhary, A.
and Agrawal, A., 2017. Deep Convolutional Neural

© 2019 Global Journals