See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/364327005

AN AUTOMATIC CLASSIFICATION METHOD FOR ENVIRONMENT FRIENDLY

WASTE SEGREGATION USING DEEP LEARNING

Conference Paper · July 2016

CITATION READS

1 1,023

1 author:

Vidhya Lakshmi
SRM Institute of Science and Technology
9 PUBLICATIONS 12 CITATIONS

SEE PROFILE

All content following this page was uploaded by Vidhya Lakshmi on 14 October 2022.

The user has requested enhancement of the downloaded file.

 2016 IEEE International Conference on Technological Innovations in ICT For Agriculture and Rural Development
(TIAR 2016)

AN AUTOMATIC CLASSIFICATION METHOD

FOR ENVIRONMENT
FRIENDLY WASTE SEGREGATION USING DEEP LEARNING

S.Sudha M.Vidhyalakshmi K.Pavithra

Assistant Professor Assistant Professor K.Sangeetha
ECE Information Technology V.Swaathi
Easwari Engineering College Anand Institute of Higher Technology Students
Chennai, India Chennai, India Information Technology,
[email protected] Anand Institute of Higher
Technology ,
Chennai, India

I. INTRODUCTION

Abstract--Recent enforcement of law by the Indian government Cities are becoming increasingly aware of the
for the welfare of sanitation workers has raised the need for an
problems related to conventional methods of waste collection.
automated system in waste management. The existing garbage
disposal system in India consists of unclassified waste collected In general, waste may be defined as unwanted materials that
from homes which are then segregated at a station manually. are not prime products which are of no further importance to
This segregation of solid waste done by manual labor can bring human in their actual form.
about many health hazards for the waste sorters in addition to
being less efficient, time consuming and not completely feasible Waste may be generated during the extraction or
due to their large amount. In our paper, we have proposed an processing of raw materials, consumption of final products
automated recognition system using Deep learning algorithm in
and human activities. They can thus be classified as industrial
Artificial Intelligence to classify objects as biodegradable and
non-biodegradable, where the system once trained with an initial waste, clinical waste and domestic waste. Improper disposal of
dataset, can identify objects real-time and classify them almost garbage has many hazards affecting all forms of life leading to
accurately. Biodegradable waste is used to generate power, contamination of air, water and soil and also causes dangerous
enrich soil and act as food to animals. This process does not harm diseases to human beings.
the earth making it valuable, ecologically safe and helps us to
protect our environment, rich ecosystem and human inhabitants The trend to instil automated waste segregation
in future.
systems into new housing development projects is rising and
Index Terms—Garbage segregation, Deep learning serves for leading the way in implementing this type of
technology. Statistics show that UK produces 330 million
tonnes of waste a year [4] and estimates show that in 2007,
480,000 fewer tons of CO2 were emitted due to separate
collection and composting of 833,000 tons of organic
materials [5]. This shows us the need for segregating and
composting garbage on a larger scale.

978-1-5090-0615-1/16/$31.00 ©2016 IEEE 65

 2016 IEEE International Conference on Technological Innovations in ICT For Agriculture and Rural Development
(TIAR 2016)

managed so as to minimize the risks to the health and safety of

II. OBECTIVE patients, the public, and the environment.

The existing garbage disposal system in India requires The economic value of waste is best known when we
manual labour for segregation of its waste. With a population segregate the waste products. Currently there is no effective
of 1.252 billion (2013) people in India, relying solely on the system for segregation of various types of wastes at a
segregation done manually is not healthy. It leads to an household level. The purpose of our project is to make a
unhealthy environment as well as can bring about many simple, low cost and user friendly segregation system for
hazards for the labourers. urban households to waste management process more
effective in India.
Our objective is to bring about an automated process to
the existing laborious method where the process is faster, V.DEEP LEARNING
cleaner and does not affect the ecosystem. The biodegradable
Deep learning is an arising field of Machine learning
products must be put to decompose and the rest, recycled. which is still in its research and mainly aims in bringing
Resources must be saved and they must not be extinguished. machine learning even more nearer to one of its focuses:
Artificial Intelligence. Machine learning concentrates on tasks
III. EXISTING SYSTEM such as recognising the images, converting speech to text,
recognition of speech and visual object, drug discovery, face
The prevailing garbage disposal system in India detection and recognition, weather forecast etc. Deep learning
consists of unorganized waste collected from habitats which techniques are used more in these applications where adaptive
are then segregated at a station. The segregation is done by learning is done. Deep learning can do cognitive learning such
manual labor which has many health mishaps for the laborers as learning the features, characteristics and attributes with the
and is time consuming and also requires financial share to the help of good algorithms which can learn by itself and deep
workers. Uncontrolled dumping of waste on outskirts of towns architecture. The family of deep learning has been increasing
and cities has created overflowing landfills which are not only which includes neural networks, various unsupervised and
impossible to reclaim because of the disorder manner of supervised learning algorithms for recognising feature such as
dumping but also has acute environmental implication in Deep Belief Network, Deep convolutional Neural Network
terms of ground water pollution and contribution to Global and Recurrent Neural Network and models which represent
warming. This has been found to reduce the average life span the probability of the hierarchy.
of the manual segregators.
Deep Structured Learning which is also termed as Deep
A new concept [1] uses a hardware component that Learning or Hierarchical Learning has risen as one of the
can sort waste at the initial stage thus making waste fastest growing areas of Machine Learning which focuses on
management more powerful and fruitful. The designed system trying to mimic all the tasks that a normal human brain can
sorts wastes into three different categories, namely metal, dry perform.
and the wet waste. A simple 8051 microcontroller forms the
heart of the system. It controls the working and timing of all It has the capacity to process complex data given as input
the sub sections so as to sort the waste into the three primary and results in giving the predicted output through effective
categories. recognition. Deep learning allows processing of multiple
2 The main disadvantage in the existing system is layers through the computational models in order to learn data
that, segregation of the waste consumes time. Size of the representations with abstraction of many layers. Deep learning
waste must be less than or equal to the dimension of the funnel uses the back propagation algorithm to discover complex
i.e. 20cm X 20cm. E-waste, Sanitary waste and medical waste structures in huge data sets and represents how a machine
cannot be segregated by the proposed system as there are should vary its inner parameters that are used to determine the
certain rules and regulations specified by government to be representation of each layer from the depiction of previous
followed for their segregation. It also uses complex hardware layer. Deep architecture has networks with multiple layers and
components for segregating the garbage. the adjacent layers are in some way connected with each other.
It is mainly used in solving the problems of representation and
IV.PROBLEM DEFINITION classification through constant learning. Perceptron which has
some weight combines all the known features for recognising
the objects.
The increasing population of India poses serious
threats like limitation of living space, education and
employment. But another serious problem is the enormous
amount of waste generated each minute by an individual.
Solid waste management is a big challenge in urban areas for
most of the countries throughout the world. An astounding
tones of waste is generated each day in India. But only 5% of
this large amount of waste is recycled. One possible solution
for this problem could be segregating the waste at the starting
stage itself. The segregation of the waste is to be properly

66
978-1-5090-0615-1/16/$31.00 ©2016 IEEE
 2016 IEEE International Conference on Technological Innovations in ICT For Agriculture and Rural Development
(TIAR 2016)

Step 1: Start

A.DEEP LEARNING IN OBJECT RECOGNITION: Step 2: The system is switched on.

Step 3: The camera takes pictures of the objects to be

identified.

Step 4: The images are analyzed with the trained data.

Step 5: The objects are then detected and localized by

boundary algorithm.

Step 6: Object identification takes place.

For each pixel in the input image, the intensity of the pixel is
encoded is the value for a corresponding neuron in the input Step 7: A prediction is made based on the identification and a
layer. For the 28×28pixel images, there are 784 neurons which probability index studied.
are given as the inputs. The output of the network can be an
input image based on the trained network‟s weights and Step 8: The prediction with highest probability is regarded as
biases.
the output.

Step 9: The object is then classified based on the output

data as biodegradable or non-biodegradable.

Step 10: Stop

Each neuron in the first hidden layer will be connected to a

small region of the input neurons. The region in the input
image is called the local receptive field for the hidden neuron
which acts as a window for the input pixels. Each connection
learns a weight by itself when trained and the hidden neuron
learns an overall bias too.

VI. PROPOSED IDEA

The proposed idea mainly focuses on the identification

and classification of the waste that is dumped in the garbage.
Usually, unsegregated waste is dumped in a landfill and made
to decay which however takes hundreds of years in the case of
non-biodegradable waste.

Our project proposes an idea where a computer on its

own is able to recognize the waste without human intervention
based on the material of the item, irrespective of its shape and
size, easily and classify them. Our proposed system can learn
by itself and thus can constantly update itself in case of new
materials. The advantages to the proposed system would Fig 1. Flow chart
include easy decomposition, lesser health hazards and faster
process that requires only an initial investment and is
automatic.

A. Algorithm VII. WORKING WITH CAFFE

67
978-1-5090-0615-1/16/$31.00 ©2016 IEEE
 2016 IEEE International Conference on Technological Innovations in ICT For Agriculture and Rural Development
(TIAR 2016)

Caffe is used with Level-Based database whose disk-key

A. Caffe Tutorial value store is chosen oven LMDB due to its interoperability to
work in UNIX, Mac, Windows and Android. The input for this
The implementation part is done using Caffe which is database is obtained from ImageNet which is managed by
a deep learning framework as software. Deep learning and Stanford and Princeton University. It is an image database
deep features have only recently achieved strong results in organized by hierarchy with each node having an average of
many tasks. Thus, a common framework and shared models 500 images. It can prove to be very useful in training a dataset
are needed to advance further research and applications. provided a high processing computer is available to process
these images.
The reasons for choosing this software includes its
comparatively: VIII. IMPLEMENTATION

High speed
Training is the first step to be practiced after the
Open source collection of dataset. Training the dataset is very important to
obtain accurate results. A large number of inputs have to be
Flexibility in all operating systems given to train the network for its own decisions.

Reliable results

Updated coding
Here for the purpose of recognizing images,
Caffe software is interoperable and can be run on
different operating systems as in Ubuntu, Windows and Mac. Train a layer of features that receive input from the
It provides a framework which can be modified for working pixels of the image.
with machine learning algorithms with a set of reference
models. It has a library with C++, Matlab and Python through The edges present in the image are recognized using
which the convolutional neural networks can be trained. It Edge detection algorithm.
needs Cuda and Gpu along which it can process around 40
million images for various tasks such as detection and Next, object parts are identified from various image
recognition. It has platforms that make developing and processing algorithms such as feature extraction.
deploying of prototypes to cloud environment easy. It has also
provisions for supporting research projects, applications in Finally, objects are identified.
industry and various applications in speech, vision and
multimedia through clear modules of code and bindings to
Python and Matlab. Improvements in deep architectures have Though the training process is described above, before
increased levels of performance of the networks. fine-tuning, a layer-by-layer pre-training is performed. Pre-
Convolutional neural networks have the capacity for back training refers to the process where the user manually trains
propagation through the convolutional layers. Caffe is an open the data by describing the inputs. This is known as supervised
framework for accessing the deep architectures. It has toolkit training where the user supervises the results of inputs
to train and deploy the networks along with examples. It also initially.
has provisions for fast implementation of the algorithms so
that they can be deployed efficiently. Using it in CPU mode In the process of backtracking, the algorithm is made to
removes the need for special hardware after training the process an image and provide an answer after which the user
models. Predefined reference models remove the need for re- provides the actual answer and a comparison is made between the
learning that may be costly and makes experimentation quick two and the algorithm „learns‟. Since there are many layers in
too. Caffe uses 4d array called blobs for storing the data. The deep learning, the output of one layer is fed as input to the
layers in Caffe can take one or more input as blobs. The layers next layer until all layers are pre-trained and ready for
can perform backward or forward pass operations. Caffe trains execution real-time.
the model with the layers in which the data layer retrieves the
images from the disk. Caffe modifies the old model weights
for developing a new task by initializing the new weights To work in GPU mode in Caffe, Cuda is required. Opencv
when needed. It can provide capacity for transferring the is installed to accelerate the GPU. BLAS performs the basic
knowledge for object detection, retrieval and recognition. End mathematical and algebraic operations. BOOST includes
to end training of network in Caffe can extract the required various libraries in C++ with pointers and few functions. The
image features. In spite of many frameworks being available, data structure of the new data layers are defined through the
the reason for choosing Caffe is that the architecture being protobuf. Gflags and Glog helps in debugging of errors
used is that of a standard one and in order to train a million through commands. Imagenet database consists of a large
images or so, it is easier. Comparatively, Torch has more collection of all possible images and so it is used as the source
functionality and is also more flexible but it needs more for training images. These images are stored in lmdb and
coding and training, thus consumes more time. leveldb of the software according to the requirement. For the
layers, a lenet_train prototxt is created. The weights for the
calculation are specified using lenet_solver prototxt. Thus the
Caffe software is trained efficiently and easily.

68
978-1-5090-0615-1/16/$31.00 ©2016 IEEE
 2016 IEEE International Conference on Technological Innovations in ICT For Agriculture and Rural Development
(TIAR 2016)

[4]http://www.slideshare.net/sushmitachanda2013/waste-disposal- system-
in-uk-and-usa
Deep Learning algorithm is implemented as the garbage
images are identified to be that of biodegradable or not. The [5]http://www.triplepundit.com/2013/01/europes-best-recycling-
steps in the identification of images is similar to that in image prevention-program
processing and edge detection and matching method are used
[6] Deep Learning by Yann Lecun, Yoshua Bengio and Geoffrey Hinton
along with grayscale as well as gradient matching methods.
Based on the outputs of these combinations of algorithms, the [7] Deep Convolution Neural Networks in Computer Vision: a Review by
image is identified. In the Caffe software, once an object is Hyeon-Joong Yoo
identified, it is made to match with the other possibilities
[8]A survey on deep learning:one small step toward AI by Dandan Mo
given during the training of data and a probability list is made
for each possibility. The image that has the highest probability [9] Basic Level Categorization Facilitates Visual Object Recognition by
index is thus confirmed to be the object. The object from the Panqu Wang and Garrison W.Cottrell
garbage is thus identified and the robotic arm will be used to
[10]An Adaptive Approach of Tamil Character Recognition Using Deep
segregate the wastes accordingly in separate mounds. Since it
Learning with Big Data-a Survey , R.Jagadeesh Kanan and S.Subramanian

[11]Automatic Detection and Classification of buried objects in GPR images

is a self learning algorithm, the system can learn and train using Genetic algorithms and Support vector machines-Edoardo Pasolli, Farid
Melgani, Massimo Donelli, Redha Attoui, Merieete De vos
by itself thus not requiring manual intervention.
[12]An introduction to Genetic Algorithms by Melanie Mitchell

[13]”e-Waste in India”, Rajya Sabha Secretriat, July 2011

IX. CONCLUSION AND RESULTS
[14]”The Gazette of India” , The prohibition of employment as Manual
Scavengers and their rehabilitation Act,2013

Keeping our environment clean and eco friendly so [15]”Genetic Algorithm for Object Recognition in a Complex Scene”, Daniel
L.Swets,Bill Punch, John Weng
that our next generation lives a disease free life is a priority.
This project aims to support that thought. The Automatic waste [16]”3D Object Recognition using Genetic
management system is a step forward to the existing system to Algorithm”TsuyoshiKawaguchi, Takeharu Baba
make the manual segregation of wastes easier. The developed
system would pioneer the work for solid waste management [17]”Artificial Neural Networks: A Tutorial”, Anil K.Jain,
process in the field of Artificial Intelligence. When properly
Jianchang Mao
trained, the system is highly efficient. The deep learning
algorithm used here has better performance with comparatively [18]”Pattern Recognition using genetic
less computation time. The emphasis of waste management algorithm”,Surapong Auwatamongkol
policies in many countries around the world has transitioned from
disposal to source segregation and recycling. The development of [19]”Object recognition using characteristic component and Genetic
Algorithm”, Pisit Phokaratkul, Sombut Foitong, Chom Kimpan
our project into a final product will help these countries in
achieving their goal. [20]”A combined pattern recognition scheme using genetic algorithm for robotic
assistance using wireless sensor networks”, Waleeh M.Alfehaid, Asad I.Khan,
This project proposal for the management of wastes is Anand Hudaya Muhamad Amin
efficient and time saving than the currently employed method [21]”A survey on genetic algorithm based Classification
where the municipality employees perform. Though this
system is simple in concept, it is very valuable and affordable. Technque for Handwritten Character Recognition”, Abhishek
Hence to ensure being automated, a system which takes lots of
Phukan, Mrinalit Borah
dataset as input without human intervention and also has the
capacity to think by itself offers the best solution. [22]”Genetic Algorithm-Neural Network: Feauture Extraction for Bio-
informatics Data”, Dong Ling Tong
It acts as an aide for reducing pollution levels and in
[23]”Feature Extraction using Gentic Algorithm”,
the long run focuses on the development of a nation and
restoration of our ecosystem. We thus conclude that our M.Pei, E.D.Goodman, W.F.Punch
project is an important asset to the society.
[24]”Facial Feature Extraction”, Gary G.Yen, Nethri
Nithianandan
REFERENCES
[25]”A survey of genetic algorithm applications for Image Enhancement and
Segmentation”, Mantas Paulinas,AndriusUsinskas
[1]”Microcontroller based Automatic Waste Segregator”,
M.K.Pushpa, Aayushi Gupta, Shariqh Mohammed Shaikh, [26]”An approach for predicting the shape and size of basic buried objects on
Stuti Jha, Suchitra V surface ground penetrating radar system” Nana Rachmana Syambas

[2]”Automated Waste Segregator”, Amrutha

Chandramohan,Joyal Mendonca,Nikhil Ravishankar, Nikhil U Baheti, Nitin
Kumar Krishnan, Suma MS

[3] “Artificial Intelligence” by Russel, Peter Norwin

69
978-1-5090-0615-1/16/$31.00 ©2016 IEEE
 2016 IEEE International Conference on Technological Innovations in ICT For Agriculture and Rural Development
(TIAR 2016)

[30]http://au.mathworks.com/videos/deep-learning-for-computer-vision-with-
[27]”Shape Recognition using Genetic Algorithms”, Ender matlab-116080.html?s_tid=gn_loc_drop

Ozcan, Chilukuri K.Mohan [31] “Basic Level Categorization Facilitates Visual Object Recognition”,
PAnqu Wang, Garrison W.Cotterel
[28]”Deep Learning of Representations:Looking Forward”, Yoshua Bengio
[32] “Deep Neural Networks for Object Detection”,Christian
[29]”Deep Learning in Neural Networks:An overview”,JurgenSchmidhuber Szegedy,Alexander Toshev, Dumitry Erha

70
978-1-5090-0615-1/16/$31.00
View publication stats
©2016 IEEE