Intelligent Based Waste Management Awareness Developed by Transfer Learning

Full text

Intelligent Based Waste Management Awareness
Developed by Transfer Learning
Krishna Mridha
Computer Engineering
Marwadi University
Rajkot, Gujarat, India
krishna.mridha108735@marwadiunive
rsity.ac.in
Abstract—Urban India produces 62 million tons of garbage
(MSW) per year, with projections of 165 million tons by 2030.
Annually, 43 million tons of urban solid wastes are collected,
with 31 million being discarded in landfills and just 11.9
million being handled. In India, many people litter the
highways as well. They probably just filled the road with
banana leaves or dried leaf bowls a few years ago; those types
of litter were not particularly dangerous because they were
environmentally friendly and could even be eaten by
abandoned animals. However, in today's India, plastic is the
most commonly littered object and cultural change is difficult
to achieve in any society. If it continues like this, India will
become an unsuitable country for living. To overcome this
crucial problem, our research will be the best solution. We
analysis on nine different classes so that the recycling authority
can recycle these easily and reduced the wastes. Our proposed
model has been designed with three different tasks. In the
beginning step, we have to capture any particular wastes. In
the second step, we have to classify these captured images with
the help of web applications that have been developed by the
Transfer Learning model (VGG16) and according to the
classification name, we have to put them into the same type
bucket so that every bucket contains only one unique type of
wastes. Finally, we have to call the authority who is involved
with recycling so that they can easily recycle them and
adequately reduce the waste. Our model accuracy is almost
95%. Even though these works are initialization stage done by
the web application but in the future, we will develop a mobile
application to make it easier and user friendly.
Keywords—garbage, dangerous, recycle, vgg16, application.
I. INTRODUCTION
India's waste management faces various problems
because of improper awareness of recycling. As a
consequence, its huge and increasing amount is affecting the
heavily populated regions. India is the world's second-most
populous country and the world's twelfth most heavily
inhabited country. Between 2015 and 2020, the expected
urban population growth rate is 5 percent. Waste
management is becoming more of a concern as the
population grows, particularly in the larger cities. Other than
that, there have been recycling techniques that can handle
segregated by gender and high moisture-low calorie waste,
which can be very useful for a country like India as
briquettes are often used in waste treatment projects.
In this context, waste management involves a wide
variety of waste bins with significant filling fluctuations
(over days, seasons, or location) and differing conditions for
emptying, ranging from occasional (a few times per week) to
very routine (daily, weekly, or monthly) (several times a
day). Other waste types (agricultural, biomedical,
pharmaceutical, mechanical, mineral, organic/inorganic, and
nuclear, to name a few) are characterized by distinct
collecting points, constant and predictable production, and
equal, usually long filling times.
Delhi is reportedly one of the most contaminated cities in
the world, according to a UNFPA survey, and one of the
problems at hand is urban waste management. Though it is
tough to control wastes, wastes can be recycled. Now, the
question will arise, how we can recycle them because wastes
are in different forms and diluted. So, to come up with a
solution, we are proposing a newer way. Our proposing
model is divided into three small different steps. In the first
step, we have to capture any particular wastes. In the second
step, we have to classify these captured images with the help
of web applications that have been developed by the Transfer
Learning model (VGG16) and according to the classification
name, we have to put them into the same type bucket so that
every bucket contains only one unique type of wastes.
Finally, we have to call the authority who are involved with
recycling so that they can easily recycle them and adequately
reduce the waste. Through, this unique way we can protect
India from serious environmental and health hazards.
II. RELATED WORK
The effective labeling and arrangement of wastes into
separate classes (such as light bulbs, paper, plastic, organic,
glass, batteries, clothes, metal, e-waste) aids in waste
utilization and disposal. Image recognition does provide
expense solutions for sustainable segregation by identifying,
classifying, and separating waste from massive piles of
garbage and trash. The classification and detection of
hazardous materials have been allowed due to
unprecedented developments in the field of image
recognition, which have provided for the use of a range of
object detection methods to identify and detect waste. Deep
learning architectures, such as deep convolutional neural
networks, are a form of deep learning algorithm, typically
resolve the disadvantages of traditional image detection
systems (based on the Haar cascade classifier, SVMs
(Support Vector Machines), or Sliding Window methods)
(DCNNs). Researchers all over the world are focused on
advances in deep learning methods for difficult object
detection tasks as a result of ground-breaking developments
in object recognition and image processing [6–10]. Different
deep learning techniques (such as CNN [11], R- family,
SSD family [12,13], YOLO [14], as well as other
classification methods (such as SVM and MLP, or Multi-
layer Perceptron) [15]) as well as other classifiers (such as
SVM and MLP, or Multi-layer Percept Researchers have
demonstrated the abilities of different deep learning
techniques in the classification and identification of rubbish
in the form of waste segregation [4,5,18,19].

The following is how the remainder of the paper is
organized: A similar analysis and its results are discussed in
Section II. Section III delves further into the methodology
and techniques used in transfer learning. Analyze the datasets
and include a performance evaluation of the outcomes in
section IV. Section V concludes by summarizing the whole
document.
III. METHODOLOGY
The key idea behind smart waste disposal is to tackle all
of the garbage while still keeping track of the entire
operation. The proposed device environment is depicted in
Figure 1. We proposed a smart waste management design to
help improve the ecosystem. A smart waste management
system includes a smart app for Android and smart
communication. The comprehensive design framework for
smart waste management can be found here.
Figure 1: Proposed Model Architecture
This paper proposed a method that integrates hardware,
software, and interaction into solutions that aims at
minimizing waste, reuse waste, and recycle waste generated
in cities through a methodology that protects India from
dangerous health risks, helps the environment and facilitates
citizenship.
In terms of education methods, this review takes a case
study method focused on real-world implementation of the
proposed approach. The final product (a real-world version
of a smart waste management app combined with database
middleware) is displayed, demonstrated, and tested. The
suggested solution is summarized below.
With this noble model, we can classify nine types of
waste materials. All classification names are light bulbs,
paper, plastic, organic, glass, batteries, clothes, metal, e-
waste. As a human, we also can classify waste but some
types of materials are very confused to identify that
particular type of waste and for that, only the recycle may be
hampered. To remove this problem we introduced the unique
model. With the help of artificial intelligence, we can easily
classify all types of waste without any confusion so that the
recycling system should be productive. As we mentioned
that every type of waste should be put into a separate bucket.
So, for keeping nine types of waste, we need nine buckets.
This will help to raise awareness for people to reduce and
recycle waste.
A. DatasetCollections and Descriptions:
The current study looks at urban wastes that are often
discarded near public areas by commuters, shoppers, and
sometimes during commercial events. There are 8514 jpg
images in this dataset. Some photos were sourced from
Google, while others were taken from a real-life scenario by
me. These photos were divided into nine categories in this
study, as shown in Table 1, including light bulbs, paper,
plastic, organic, glass, batteries, clothes, metal, e-waste.
Besides, Figure 2 depicts a detailed description of the waste
items allocated to these specified classes, as well as the
number of quantity and picture, as well as their class
distribution.
Table 1: Sample images with their associated classes are
depicted.
Class 0 (Battery)
Class 1(Cloth)
Class 2 (E-waste)
Class 3 (Glass)
Class 4(Light-bulb)
Class 5 (Metal)
Class 6 (Organic)
Class 7(Paper)
Class 8(Plastic)
Table 2: Class Items with Numbers of Item Quantity
Class
Item
Quantity
Battery
Small sealed lead-acid
batteries, nickel_cadmium
_batteries,
lithium_ion_batteries
1222
Cloth
chiffon, cotton, Crepe,
Denim, Lace, Leather, Linen,
Satin.
729
E-waste
ICT and Telecommunications
624

equipment, office electronics
Glass
Float Glass, Toughened
Glass, Tinted Glass,
Obscured Glass.
773
Light-bulb
Fluorescent Lamps, Compact
Fluorescent Lamps (CFL),
Halogen Lamps, Light
Emitting Diode (LED)
651
Metal
Ferrous Metals, Copper,
Brass, Tin, Lead, Bronze,
Zinc
1092
Organic
vegetables, fruits, fast food,
bread, rice, grass, leaves
671
Paper
white office paper,
newspaper, colored office
paper, cardboard, white
computer paper, magazines,
catalogs, and phone books.
1468
Plastic
Polyethylene Terephthalate,
High-Density Polyethylene,
Polyvinyl Chloride, Low-
Density Polyethylene,
Polypropylene, Polystyrene
Miscellaneous Plastics.
1241
Total = 8417
Figure 2: Statistical Analysis of Classes
B. Data Augmentation:
A lack of data makes it difficult to use deep learning
models like convolutional neural networks. In certain cases,
imbalanced classes may be a further stumbling block;
although there might be enough data for certain classes,
under-sampled classes may suffer from low class-specific
accuracy. This is a natural phenomenon. When a model is
trained on only a few instances of a class, it is less likely to
predict class invalidation and test applications.
Image augmentation was designed to produce training
data from an existing collection to avoid the high cost of
collecting hundreds of training images. Image augmentation
is the method of editing images already in a training dataset
to produce several altered versions of the same image. This
not only gives us more photos to train on, but it also exposes
our classifier to a broader range of lighting and color
conditions, making it more robust.
However, before using any technique, first, we have to
resize the images. The dimensions of the images gathered
from the web can vary. Since most neural networks have
fully connected layers, the images fed to the network must
be of a specific size. As a result, before we begin image
augmentation, let us first preprocess the images to the size
that our network requires. We get the advantages of filtering
them in pipelines with reasonable quantities.
To begin, we must first construct an image generator by
calling the ImageDataGenerator () function and passing it a
set of examining the changes we want to be made to the
picture. We'll then use our image generator's fit () tool to
add the improvements to the photos batch by the group. By
default, the changes are applied at variance, which means
that not every picture will be modified each time. Should
you wish, we can also use Keras. preprocessing to distribute
enhanced image files to a folder to create a large dataset of
modified photos. We can pass more parameters in
ImageDataGenerator () function. For instance, Scaling,
Translation, Rotation (at 90 degrees), Rotation (at finer
angles), Flipping, Adding Salt and Pepper noise, lighting
condition Perspective transform. The motive of using these
extra parameters to increasing the numbers of training
images so that our desired accuracy will be stable.
C. Transfer Learning
In machine learning methods, supervised learning will
fail to create correct classification models due to a lack of
labeled data. To fix the lack of labeling problem, transfer
learning was developed [1]. It aims to improve learning
performance by transferring data from multiple source
domains to a single target domain [2]. With just a few
labeled training images, image recognition, for example, can
be modeled as a target learning exercise. Fortunately, such
texts relating to pictures, such as picture explanations or
marks around photographs, may be captured and used to
classify images in a target domain using text data (a source
domain). In transfer learning, the source and target domains
often have function vector spaces. A comparison of classic
machine learning and knowledge transfer approaches is
shown in Figure 2.
Figure 3: Basic ML Approach vs Transfer Learning
Approach

When opposed to other approaches, convolutional neural
networks are superior in image recognition [3]. Even if
they're not invariant to scale and fermentations, pre-trained
CNN models trained on a broad image dataset (i.e.
ImageNet) will produce an image features map vector to
rotating and shape changes [16]. Any of these deep models
are open-source, easy to use, and modify.
Figure 4: The VGG16 Model consists of 16
Convolutional and Max Pooling layers, 3 Dense layers
for the Fully-Connected layer, and a 1,000-node output
layer.
Now, the process of pre-defined model VGG16 is
established by the Transfer Learning way. For Transfer
Learning we do not require the fully connected layer (Dense
Layer) so that we have to remove the dense layer and have
connected with VGG16. The ImageNet data set has 1000
categories but in our case, we have only two labels. So we
have to develop our output as 2 labels.
1. Pre-trained weights from a learning algorithm on a
large dataset are loaded in.
2. Freeze all weights in the lower (convolutional)
layers: the layers to freeze are determined by how
closely the current task resembles the original
dataset.
3. Replace the network's upper layers with a custom
classifier with the same number of outputs as
classes.
4. Just train the custom classifier layers for the
mission, allowing the model to be optimized for
smaller datasets.
IV. RESULTS AND DISCUSSION
In this result section, we are showing our desired
accuracy against training and testing images. Even though
we have gotten 95% accuracy, it is the best accuracy yet.
Because there is no such type of research is happed till no.
All types of classes are made by me and also I added more
class images as well as developed a web application. With
the help of a web application, users can take the images of
the waste and upload these images for classification. After
knowing this waste type, the users can keep this waste in a
particular waste bucket. And finally, they will call the
authority for collecting these wastes so that the authority can
be recycled these waste.
The motive of this paper is to create awareness about the
curse of waste among the peoples in India. India is going to
fall into a huge environmental problem after few years. To
overcome this crucial problem, we have developed this
unique application. This application is very user-friendly
and less complex. This application will help a lot to manage
the wastes and also help to collaborate with recycle
authority. By this application, the user can request the
nearest authority to collect the wastes by submitting an
online form. The online form should be collect the user
information related to address, contact number, and so on.
In this paragraph, we are going to show the user
interface of our developing application and we will mention
the life cycle of this application also. The way of using this
application is mentioned below.
• First, take the images by clicking the “Classify
your waste material” button (Red Arrow).
Figure 5: First step: Taking Image
• After taking images, the machine automatically
showing the class name in the right block of the
same page. There will be showing the class name
as well as more details about the particular class.
Figure 6: Second Step: Classify the Image and see the
output

• The last step is according to class, we have to keep
this particular waste into the same particular
bucket.
Figure 7: Third step: Keep waste to the particular
bucket
V. CONCLUSION AND FUTURE SCOPE
An intelligent-based waste management system was
a dream for everyone. But now with the help of Artificial
Intelligent, it has become a narrow work. The field of
Artificial Intelligent is very wide and stable. Our developing
model is also a part of AI. This is the very fast starting of
our research. This is the very first step to increase the
awareness of waste and in the future, we can work deeply.
Now, our proposed model is implemented by Transfer
Learning Method (VGG16) and deployed as a web
application and we have a plan to available in Android and
IOS operating systems so that from low society people to
high society people can take this opportunity. The very first
time, we want to work with the municipal area so that we
can realize our drawbacks. If we face any difficulties then
we may overcome all of the difficulties and after that, we
can start the project with full force all over India.
REFERENCES
[1] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE
Transactions on Knowledge and Data Engineering, vol. 22, no. 10.
pp. 1345–1359, 2010.
[2] F. Zhuang et al., “A Comprehensive Survey on Transfer Learning,”
Nov. 2019
[3] J. Wang, Y. Yang, J. Mao, Z. Huang, C. Huang, and W. Xu,
“CNNRNN: A Unified Framework for Multi-label Image
Classification,” in Proceedings of the IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, 2016, vol.
2016-Decem, pp. 2285–2294.
[4] Wang, Y.; Zhang, X. Autonomous garbage detection for intelligent
urban management. MATEC Web Conf. 2018, 232, 01056.
[5] Devi, R.S.S.; Vijaykumar, V.R.; Muthumeena, M.Waste segregation
using a deep learning algorithm. Int. J. Innov. Technol. Explore. Eng.
2018, 8, 401–403.
[6] Lu, H.; Zhang, M.; Xu, X.; Li, Y.; Shen, H.T. Deep Fuzzy Hashing
Network for Efficient Image Retrieval. IEEE Trans. Fuzzy Syst.
2020.
[7] Lu, H.; Li, Y.; Chen, M.; Kim, H.; Serikawa, S. Brain Intelligence:
Go beyond Artificial Intelligence. Mob. Netw. Appl. 2018, 23, 368–
375
[8] Lu, H.; Li, Y.; Mu, S.; Wang, D.; Kim, H.; Serikawa, S. Motor
anomaly detection for unmanned aerial vehicles using reinforcement
learning. IEEE Internet Things J. 2018, 5, 2315–2322
[9] Chen, Z.; Lu, H.; Tian, S.; Qiu, J.; Kamiya, T.; Serikawa, S.; Xu, L.
Construction of a Hierarchical Feature Enhancement Network and Its
Application in Fault Recognition. IEEE Trans. Ind. Inform. 2020.
[10] Huang, R.; Gu, J.; Sun, X.; Hou, Y.; Uddin, S. A Rapid Recognition
Method for Electronic Components Based on the Improved YOLO-
V3 Network. Electronics 2019, 8, 825
[11] Dundar, A.; Jin, J.; Martini, B.; Culurciello, E. Embedded streaming
deep neural networks accelerator with applications. IEEE Trans.
Neural Netw. Learn. Syst. 2017, 28, 1572–1583.
[12] Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.Y.;
Berg, A.C. SSD: Single shot multibox detector. In Proceedings of the
Lecture Notes in Computer Science (including subseries Lecture
Notes in Artificial Intelligence and Lecture Notes in Bioinformatics);
Springer: Berlin/Heidelberg, Germany, 2016; Volume 9905 LNCS,
pp. 21–37.
[13] Fu, C.-Y.; Liu, W.; Ranga, A.; Tyagi, A.; Berg, A.C. DSSD:
Deconvolutional Single Shot Detector. arXiv 2017,
arXiv:1701.06659.
[14] Lu, S.; Wang, B.; Wang, H.; Chen, L.; Linjian, M.; Zhang, X. A real-
time object detection algorithm for video. Comput. Electr. Eng. 2019,
77, 398–408.
[15] Dollár, P.; Wojek, C.; Schiele, B.; Perona, P. Pedestrian detection: An
evaluation of the state of the art. IEEE Trans. Pattern Anal. Mach.
Intell. 2012, 34, 743–761
[16] A. S. Tarawneh, C. Celik, A. B. Hassanat, and D. Chetverikov,
“Detailed Investigation of Deep Features with Sparse Representation
and Dimensionality Reduction in CBIR: A Comparative Study,” Nov.
2018
[17] Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol.
521, no. 7553. pp. 436–444, 2015.