Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)

IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

Deep Learning based Smart Garbage

Classifier for Effective Waste Management
Sidharth R, Rohit P, Vishagan S, Karthika R, Ganesan M
Department of Electronics and Communication Engineering,
Amrita School of Engineering, Coimbatore
Amrita Vishwa Vidyapeetham, India-641112
[email protected] m, [email protected], [email protected]

Abstract- India is the second most populated country in sensors and electrical circuit. Whereas in some
the world and it is still facing hindrances to its methods the number of the classification groups were
development on waste management. It is believed that limited .Most of the techniques either use highly
10 million tons of waste is produced just by the
priced equipment’s or they are not practical for daily
metropolitan cities in India. In this work, a way to
classify the waste and find the category of it is proposed
purposes.
with a well-defined and labelled data set of images
With the help of this method the output can be
consisting of categories (plastic, paper, cardboard,
implemented on a futuristic automation model and it
metals) using Convolutional Neural Network (CNN).
Images are categorized based on their properties by the
can be useful in every household to segregate the
help of a self-learning neural network. The designed waste. Four categories of image waste have been used
classifier learns from the image data provided for namely cardboard, metal, paper and plastic. Metal and
training purpose. The classifier uses the method of plastic are harsh substances to the nature therefore, it
supervised learning where the algorithm learns from a must be recycled and used effectively. Efficient
labelled data set. With this method a testing accuracy of method to sort the waste in developing countries and
76% is achieved. developed nations is a major concern. Importance
Keywords-Convolutional Neural Network, Waste must be given for the effective management of waste
Segregation, Tensorflow. as it may affect the economy of the nation in the
longer run [1]. About lakhs of metric tonne of solid
I. Introduction waste is piled up every day so it is impossible for
human labour to classify and sort the recyclable waste
Waste classification plays an important role in waste separately hence this classifier technique helps to
management. It has been estimated that 15,000 metric segregate the solid waste safely and quickly in a large
ton of solid waste is generated per day, which almost scale process if it’s been automated.
results in 55 lakh metric ton per year, which results in
II. Literature survey
severe global crisis. Hence there is an urge in need of
treating the waste and the only way to treat the waste Zol Bahri et al., [2] focused on separating plastic
is by recycling the waste. So, segregating waste is objects and paper objects from the waste. They have
very useful as it increases the rate of recycling and used a camera to sense the image and conveyer belt to
biodegradable waste can be effectively used. Manual move the waste and drop it in different bins of each
waste segregation can be a very hazardous, tedious category. Mahmudul et al., [3] developed a system to
and time-consuming process hence there is a need for classify the waste materials as metal, glass and
an alternative method to segregate waste. transparent. This method uses a metal sensor, glass
sensor, LDR (light dependent resistor), LASER, IR
The alternate approach must be time efficient
and cost effective .A small error in the classification transmitter, receiver and a microcontroller.Servo
motor is used to move the trash to the respective bins.
can lead to various serious environmental damages.
Therefore deep learning based segregation is chosen. George E. Sakr et al., [4] compared two methods,
they explored deep learning and support vector
This is a safer and secure approach. The proposed
classifier uses a Convolutional Neural Network machine (SVM) to classify the waste. They concluded
that the support vector machine outperformed the
(CNN) which extracts the properties of the images in
alexnet. Cenk Bircanoglu etal., [5] compared on
the dataset to identify the category of waste.
However, the technology is so rapidly developing several different deep Convolutional neural network
architectures as well as different Optimization
hence there have been many techniques that are
rapidly developed and tested. In the past, number of methodologies. The model they used were Mobilenet,
Recyclenet,ResNet50, Inception, ResNet, Xception,
techniques were proposed which includes complex

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1086

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on April 08,2022 at 01:56:36 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)
IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

Densenet121, Densenet169 and Densenet201. The

training samples were limited. Yijian Liuin et al., [6]
demonstrated which consist of a hardware module of
raspberry pi, SURF-BoW algorithm and multi-class
SVM classifier. They have used 5 categories such as
batteries, bottles, cans, paper-balls and paper-boxes.
Olugboja Adedeji et al., [7] demonstrated an
intelligent system for waste classification using pre-
trained residual convolutional neural network Yin
Shen et al., [8] proposed a model which classifies
impurities of wheat using Convolution Neural
Network.Dataset of normal wheat and five impurities
were collected and applied an image processing
Fig.1. Sequential flow of the Convolutional Neural Network [11]
technique with Wiener filtering algorithm and the
Multi-scale retinex enhancement algorithm. They The images are initially converted to a pickle file.
trained, sample dataset by rotating the sample image Pickling an image is converting the python objects
at seven angles, horizontal and vertical flip of the into a byte stream to store it in a file .This file is used
sample in different directions as well. Kancharla as the input image to the neural network. Convolution
Tarun et al., [9] designed a model which sorts plastic is the multiplication of the matrix of image pixel and
and non-plastic waste using Convolutional Neural a filter matrix. These matrices consist of pixels of the
Network. They used RFID attached to the material at images in the data sets. This convolution is done for
a time of manufacturing to segregate the type of all the pixels. And use conv2d to give input images as
a 2-Dimensional matrix. In our classifier classifying
waste. They designed a prototype which preform as
the images in four categories as cardboard, metal,
real-time machine. paper and plastic is done
III. Methodol ogy The output layer consist of dense and the activation
used is softmax with 4 different classes. Fig 2 depicts
CNN is deep learning algorithm which has been the graph of a softmax function [12].
inspired from visual cortex of human beings. Deep
learning technique is used in various fields such as in
biomedical advancement [10], image classification
[11], and self-driving cars and so on. The ConvNet
needs less amount of pre-processing as compared to
other classification algorithms. It is a fully connected
algorithm which means each neuron in one layer is
connected to all other neuron in all layers.so CNN is a
self-learning algorithm which transfers the image into
matrix and passes it to filter, then convolution is
performed to this matrix and the result is stored as a
model. The effectiveness of a CNN is determined by
the quality and quantity of images provided in the
datasets for the training. Here, in this model, a
supervised learning method is adopted where the data
Fig .2. Graph of a softmax function [12]
set is well defined and labelled. CNN is a simple and
more effective method of deep learning as compared
to various other methods. ( ) ( )
In fig.1 the functional sequence of convolutional ∑
neural network is depicted [11].

The softmax ensures that the values of the output

node is in the range of 0 to 1.Simply softmax
regression is nothing but a similar to a logistic
regression which normalizes the given value to vector
which follows a probability distribution that sums up
to one.

In the equation (1) computed the exponential of the

input parameter and the sum of the exponential

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1087

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on April 08,2022 at 01:56:36 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)
IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

parameters of other inputs ∑ the softmax is

the ratio of the input parameter needed to the ratio of
the all input parameter. z is the sum of multiplication
of all the weight matrix and the feature matrix.

In fig. 3. The process of the Convolutional Neural

Network is shown.

Fig. 4. Sample images

A validation set of 0.1 i.e. 10% of the total images in

the dataset is used for the validation purpose therefore
in total there are 1,889 trained images and 188 tested
images. The distribution of the images of dataset is
listed in Table 1.
T ABLE I. DIST RIBUT ION OF DAT ASET

Category Training Validation Image

Images Images Size
Cardboard 403 40 50x50

Metal 410 41 50x50

Paper 594 59 50x50

Plastic 482 48 50x50

4.2 Results and Discussion

The Training and Validation graph of our approach is

Fig. 3. Design of the ConvNet
given in fig. 5. The following results of accuracy
IV. Results were obtained in our experiment.

4.1 Dataset

The dataset that used is trashnet dataset [13]. The

dataset consist of 6 categories total of 2527
images but only four essential categories is used
and the total images of 2077. The sample images
are given in fig. 4. and the image distribution of
the data set is as follows:
 Paper: 653 images
 Metal:451images Fig. 5. T raining and Validation graph for accuracy
 Plastic:530 images
 Cardboard:443 images A maximum accuracy of 76% at 100 Epochs and
image size of 50x50 is achieved.
TABLE II. ACCURACY VS EPOCH

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1088

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on April 08,2022 at 01:56:36 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Fifth International Conference on Communication and Electronics Systems (ICCES 2020)
IEEE Conference Record # 48766; IEEE Xplore ISBN: 978-1-7281-5371-1

Epochs Accuracy (%) [6] Liu, Yijian, King-Chi Fung, Wenqian Ding, Hongfei Guo, T ing
Qu, and Cong Xiao. "Novel Smart Waste Sorting System based on
Training Testing Image Processing Algorithms: SURF-BoW and Multi-class
SVM." Computer and Information Science 11, no. 3 (2018): 35-49
.[7] Adedeji, Olugboja, and Zenghui Wang. "Intelligent Waste
50 97.47 73.54 Classification System Using Deep Learning Convolutional Neural
Network." Procedia Manufacturing 35 (2019): 607 -612.
100 99.12 76.19 [8] Shen, Yin, Yanxin Yin, Chunjiang Zhao, Bin Li, Jun Wang,
Guanglin Li, and Ziqiang Zhang. "Image Recognition Method
150 99.94 74.60 Based on an Improved Convolutional Neural Network to Detect
Impurities in Wheat." IEEE Access 7 (2019): 162206 -162218.
200 99.29 72.49 [9] T arun, Kancharla, K. Sreelakshmi, and K. P. Peeyush.
"Segregation of Plastic and Non-plastic Waste using Convolutional
Neural Network." In IOP Conference Series: Materials Science and
Engineering, vol. 561, no. 1, p. 012113. IOP Publishing, 2019.
The training and testing accuracy were 99.12% and [10] Rajkumar, A., M. Ganesan, and R. Lavanya. "Arrhythmia
76.19% respectively for 100 epoch as shown in fig 6. classification on ECG using Deep Learning." In 2019 5th
As the test data is unseen by our model, and lot of International Conference on Advanced Computing &
Communication Systems (ICACCS), pp. 365-369. IEEE, 2019.
parameters are considered the testing accuracy is less [11] Hacibeyoglu, Mehmet. (2018). “ Human Gender Prediction on
than our training accuracy. In this model, able to Facial Mobil Images using Convolutional Neural Networks”,
classify a group of waste category efficiently and cost International Journal of Intelligent Systems and Applications in
Engineering. 3. 203-208. 10.18201/ijisae.2018644778.
effectively which is not addressed in the earlier [12] Shen, Leixian, Qingyun Zhang, Guoxu Cao, and He Xu. "Fall
approaches. This model is an easier approach to the detection system based on deep learning and image processing in
classification of the waste. In the earlier approach cloud environment." In Conference on Complex, Intelligent, and
Software Intensive Systems, pp. 590-598. Springer, Cham, 2018.
categories were limited but in this classifier model a [13] T hung, G. "T rashnet." GitHub repository (2016).
lot of categories can be implemented for the
classification purpose.

V. Conclusion

This classifier is able to classify accurately and

recognise different categories . The less accuracy is
caused when classifying the metal from other
categories. The limited data set is the main drawback
which causes a limited accuracy. This classifier can
be further developed with a better dataset with a good
quality and quantity of images . Hereby, in the future
improve the proposed model with more accuracy by
training a dataset with more number of images and
finding other alternatives for improving the accuracy.
Soon it can be brought with automation for the
practical applications in every household.

REFERENCES
[1] Read, Adam D. "“A weekly doorstep recycling collection, I had
no idea we could!”: overcoming the local barriers to participation."
Resources, Conservation and Recycling 26, no. 3-4 (1999): 217-
249.
[2] Razali, Zol Bahri, and S. Madasamy. "An Application of Image
Processing for Automated Mixed Household Waste Sorting
System." In International Conference on Man-Machine System
(ICoMMS2012), Penang, vol. 27, p. 28. 2012.
[3] Russel, Mahmudul Hasan, Mehdi Hasan Chowdhury, Md
Shekh Naim Uddin, Ashif Newaz, and Md Mehdi Masud T alukder.
"Development of automatic smart waste sorter machine." In
International Conference on Mechanical, Industrial and Materials
Engineering, vol. 1. 2013.
[4] Sakr, George E., Maria Mokbel, Ahmad Darwich, Mia Nasr
Khneisser, and Ali Hadi. "Comparing deep learning and support
vector machines for autonomous waste sorting." In 2016 IEEE
International Multidisciplinary Conference on Engineering
T echnology (IMCET ), pp. 207-212. IEEE, 2016.
[5] Bircanoğlu, Cenk, Meltem Atay, Fuat Beşer, Özgün Genç, and
Merve Ayyüce Kızrak. "RecycleNet: Intelligent waste sorting using
deep neural networks." In 2018 Innovations in Intelligent Systems
and Applications (INIST A), pp. 1-7. IEEE, 2018.

978-1-7281-5371-1/20/$31.00 ©2020 IEEE 1089

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on April 08,2022 at 01:56:36 UTC from IEEE Xplore. Restrictions apply.