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

net/publication/339599806

Flower Identification Using Machine Learning This Report conferred to the

Department of CSE of Daffodil International

Thesis · December 2018

CITATIONS READS

0 5,446

1 author:

Md. Mizanur Rahman Nayeem

Daffodil International University
1 PUBLICATION   0 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Flower Identification Using Machine Learning View project

All content following this page was uploaded by Md. Mizanur Rahman Nayeem on 03 March 2021.

The user has requested enhancement of the downloaded file.

 Flower Identification Using Machine Learning
BY

Md. Mizanur Rahman

ID: 151-15-4910

Akash Ahmed Khan

ID: 151-15-4883

Md. Baher Uddin Shameem

ID: 151-15-5299

This Report conferred to the Department of CSE of Daffodil International

University in Partial Fulfillment of the Requirements for the Degree of
Bachelor of Science in Computer Science and Engineering

Supervised By

Dr. Syed Akhter Hossain

Professor and Head
Department of Computer Science and Engineering
Daffodil International University

DAFFODIL INTERNATIONAL UNIVERSITY

DHAKA, BANGLADESH
DECEMBER 2018
©Daffodil International University ii
©Daffodil International University iii
 ACKNOWLEDGEMENT

We take this opportunity to express our sincere gratitude to Prof. Dr. Syed Akhter
Hossain Sir, Department of CSE Daffodil International University, Dhaka. Deep Knowledge
& keen interest of our supervisor in the field of “Machine Learning” to carry out this project.
His endless patience, scholarly guidance, strong motivation, continual encouragement,
constant and energetic supervision, constructive criticism , valuable advice, reading many
inferior draft and correcting them at all stage have made it possible to complete this project.

We would like to express our heartiest gratitude to the Almighty Allah and also to other faculty
member and the staff of CSE department of Daffodil International University to finish our
project.

We would like to thank our entire course mate in Daffodil International University, who took
part in this discuss while completing the course work.

Finally, we must acknowledge with due respect the constant support and patients of our
parents.

©Daffodil International University iv

 ABSTRACT

Beauty is incomplete without flower. Bangladesh is the land of flower. In our everyday life,
On the way of walking, beside the rail line or in our garden we used to see a lot of flower. But
in most case we have no knowledge about that flower. Even we don’t know its name. In that
case we choose this idea to research and develop our project. That will introduce people about
that unknown flower which they see but don't know about that. Our developed application
recognizes the flower in real time by using mobile camera. This project is an attempt at using
the concepts of neural networks to create an image classifier by Tensorflow on Android
platform. Convolutional neural networks are a popular realm of machine learning, and are
widely used in image classification. So that we choose this topic to research about image
classification by CNN and Tensorflow. Though we developed an application of our research
project but we have a lot of plan to do more research to upgrade our system.

©Daffodil International University v

 TABLE OF CONTENTS

CONTENTS PAGE
Acknowledgement iv
Abstract v
CHAPTER
CHAPTER 1: INTRODUCTION 1-4
1.1 Introduction 1
1.2 Motivation 2
1.3 Rationale of the Study 2
1.4 Research Questions 3
1.5 Expected Outcome 4
CHAPTER 2: BACKGROUND 5-9
2.1 Introduction 5
2.2 Related Works 5-6
2.3 Research Summary 7
2.4 Scope of the Project 8
2.5 Challenges 9
CHAPTER 3: RESEARCH METHODOLOGY 10-18
AND IMPLEMENTATION
3.1 Introduction to Machine Learning 10
3.2 Convolutional Neural Network 11
3.3 Design of Dataset 12
3.4 Data Collection Procedure 13
3.4 Statistics of Data 14
3.5 Implementation Requirements 14
3.6 Implementation of Network Design 15-17
3.7 Design Development 18

©Daffodil International University vi

CHAPTER 4: EXPERIMENTAL RESULTS AND 19-27
DISCUSSION
4.1 Introduction 19
4.2 Experimental Results 20-21
4.3 Main Graph 22
4.4 Descriptive Analysis 23-27
4.4 Experiment Summary 27

CHAPTER 5: SUMMARY, CONCLUSION,

RECOMMENDATION AND IMPLICATION FOR FUTURE
RESEARCH
5.1 Summary of the Study 28
5.2 Future work 29
5.3 Conclusions 29

References 30
Appendices: Accuracy Value For Some Training Steps 31-33

©Daffodil International University vii

 List OF FIGURES

FIGURES PAGE

Figure 1.5 : Rose Identification with highest confidence value 4

Figure 2.2 : PictureThis application feature graphic image 6 6
Figure 2.3 : PlantSnap application feature graphic image 6
Figure 3.1: Learning System Structure 10
Figure 3.2: Typical CNN design 11
Figure 3.3: Design of Dataset 12
Figure 3.4: Dataset of Shapla 13
Figure 4.1: Image pre processing by tensorflow 19
Figure 4.2: Training Step 20
Figure 4.3: Experiment Result using Android Application 21
Figure 4.4: Main Graph of Tensorflow Computation 22
Figure 4.5: Train Accuracy Graph 23
Figure 4.6 : Cross Entropy Graph 24
Figure 4.7: Final Training Summaries Graph 24
Figure 4.8: Activation Graph 25
Figure 4.9: Final Training Summaries Histogram 25
Figure 4.10: Final Training Summaries Histogram 26
Figure 4.11 : Train Accuracy Value 27
Figure 4.12: Cross Entropy Value 27

LIST OF TABLE
PAGE
TABLE
Table 3.4 : Statistics of Data 14

©Daffodil International University viii

 CHAPTER 1

1.1 INTRODUCTION
Flower is a very important part of nature. Mostly we identify a plant through its flower.
Experienced botanists do this identification of flower but a naive person will have to consult flower
guidebooks or browse any relevant web pages on the Internet through keywords searching. Our
system can recognizes the flower in real time using mobile camera. Currently this Android app
can identify around 10 flowers. Most important thing is that this app can fully works in offline.
We are continuously working to add more flowers to identify. Everyday we see a huge number of
flower species in our house, parks, roadsides, in farms, on our rooftop but we have no knowledge
of that flower species or their origin. Even we have no idea about its name. There are several
guidebooks for flowers knowledge but it becomes quite difficult to find the name when have the
picture. Even the Internet sometimes is not useful. But it is quite difficult for human brain to
memorize all the species they see. Even some flower is similar to look at. This application
recognizes the flower in real time by using mobile camera.

The purpose of this project is to use Tensorflow, an open-source dataflow and machine learning
library, to build an image classifying Convolutional Neural Network (CNN) for classifying the
flower image. Tensorflow, in addition to providing developers a simple way to build neural
network layers, can also be run on mobile platforms such as Android. The ultimate goal of this
project is to design and optimize a convolutional neural network for use with flower classification,
and eventually build a simple classification app for mobile devices around the trained network.
The mobile app will allow users to try and classify flowers while outdoors or offline [1].

©Daffodil International University 1

1.2 MOTIVATION
There have a special motivation behind this research and project work. Me and my cousin used to
walk along the railway line in afternoon. That time we used to observed a lot of unknown flower
around the area. We were curious about that flowers but we could not recognize it. That time we
discussed about an idea to make a system which can identify our desired flower automatically.
Since then our journey has started and dream have come true now. We developed an android
application which can introduce people about that unknown flower which they see but don't
identify that. Hope our research & developed project will be helpful for the curious people.

1.3 RATIONALE OF THE STUDY

At the primary stage of our journey looked for research paper and tools related with our idea. Even
we deeply searched the Internet to find if there any system that can identify flower automatically.
But unfortunately we found very few resources. There exists a handful papers in there. That time
we take this as a challenge and start our research to make a system which can recognize flower in
real time. After a lot of hard work our project now come to light finally. Now our application can
identify around 10 Bangladeshi common flowers with some foreign flowers also. The most unique
feature of project is it can identify flower in real time. For some kind of flower it provides 100%
accuracy rate. Different flower is same to look at we know. If this occur during identification time
that time our application shows 3 or 2 most similar result. Convolutional Neural Network (CNN)
and Image Classification technique used in our project which can identify flower with the
confidence level from 0 to 1. We use Tensorflow an open source library for training process.

©Daffodil International University 2

1.4 RESEARCH QUESTIONS
1. Can a machine identify flower before the human eye?
2. Can a system recognize different color of same flower using machine learning? Ex: Pink
Rose / White Rose
3. Is convolutional neural network (CNN) technique is perfect to identify flower accurately?

©Daffodil International University 3

1.5 EXPECTED OUTCOME
1. Identify flower in real time which include native Bangladeshi flowers along with some foreign
flowers also.

2. Produce a convolutional neural network which is capable correctly classifying images of flowers
with an average confidence level of 95% or more. Sometimes it provides 100% accuracy rate for
some kind of flowers.
3. Collected a dataset of over 5000 images of flowers using their genus-species classification as
the Google Image search term.

The following figure showing the output of the application which identify Rose with highest
confidence value by using mobile camera.

Figure 1.5 : Rose Identification with highest confidence value

©Daffodil International University 4

 CHAPTER 2
BACKGROUND

2.1 INTRODUCTION
Since the invention of computer, it has begun to change our daily life. It improves the quality of
our lives. An impressive idea is to let a computer think and learn as like human being. Basically
machine learning developed skills to develop the computer itself with the knowledge given. There
are many applications of CNNs, one of the most popular of which is image classification, and
CNNs have even been used specifically for flower classification. These flower classification neural
networks, however generally use datasets with very specific details. Thus, this project is
explorative and aimed at learning how to design a neural network using Tensorflow, but ultimately
has practical applications for developers, botanists, or nature enthusiasts. The dataset used for
training was collected through Google Images, and the images of the plants themselves are very
general, rather than of specific characteristics of the plants, and typically include most of the
identifying characteristics of a flower [3].

2.2 RELATED WORKS

Convolutional Neural Network (CNN) and Image Classification is the most used technique in
machine learning. On the other hand Tensorflow an open source library developed by Google
Brain team is the hot topic nowadays in the field of machine learning. There are some good
publications of Tensorflow. But It is very little compared to the importance. In the following
overview of the related works, we try to cover the works that are mostly related to ours to our best
knowledge. Image Classification by Tensorflow is a relatively new topic. But the idea of reliable
classification has made it increasingly important [1].

Recent history of modern software technology we can see lots of use of machine learning and
image processing in our day to life. Like Police using face detection technology to identify
criminal. Facebook using AI technology to match your face ID with any photo of you that has been
uploaded. Pinterest use this technology to find similar photos. Traffic police also find out vehicles
by using this technology. There is so many big projects on this topic, those are given below,

©Daffodil International University 5

 » PICTURETHIS

Figure 2.2 : PictureThis application feature graphic image

Among the popular software This is a Plant identification software developed by “Glority
Software limited” running on both Android and IOS. This Software let you identify plant
by taking a picture with 90% accuracy rate less than a second.

» PLANTSNAP

Figure 2.3 : PlantSnap application feature graphic image

This is also a popular software like PictureThis, developed by PlantSnap INC. This
software let us to identify plants, flowers, cacti, mushrooms etc. by taking a picture
of that plant. There are 585,000 species of plants and trees in their database. So their
accuracy rate is 90% accurate.

©Daffodil International University 6

2.3 RESEARCH SUMMARY
The purpose of this project is to use Tensorflow, an open-source dataflow and machine learning
library, to build an image classifying Convolutional Neural Network (CNN) for classifying the
flower image. Tensorflow, in addition to providing developers a simple way to build neural
network layers, can also be run on mobile platforms such as Android. The ultimate goal of this
project is to design and optimize a convolutional neural network for use with flower classification,
and eventually build a simple classification app for mobile devices around the trained network.
The mobile app will allow users to try and classify flowers while outdoors or offline. After
examining, inquiring, researching and comparing all the related works on Flower Identification,
we come to have different and many types of conclusion. First of all, none of them are using real
time object detection and identification. All of them are just stuck into by taking picture of object
and search through their database to identify the object. Secondly is accuracy rate, most of the big
project hit the accuracy rate up to 90% so they say. But the community feedback shows the
different color. Users complain about the miss detection of plants and less information about it. It
also take long time for analyzing and searching their database. Because of our real time flower
identification, it is more like instant detection of flower.

©Daffodil International University 7

2.4 SCOPE OF THE PROJECT

Developers
The project is complete ready for practical use, someone who wishes to learn the basics of CNNs
and Tensorflow, substituting their own classes. The project is essential for anyone that wants to
improve their project. It have scope to improve both the dataset and the CNN architecture,
however, this will require more research and knowledge of CNN design. Additionally, someone
who is interested in Tensorflow and Android development or front end development can be
benefited from this work.

Scientist or Botanist
This research and project will be helpful for botanists or biologists, although any scientist would
likely wish to use their unique dataset. The project would most likely be useful to scientists for
classifying large amounts of data and generating statistics for the dataset. Using our Android
application Botanists or other scientists dealing with flowers to identify easily.

Traveler or Nature Enthusiasts

Our developed project will be much helpful while spending time outdoors to identify flowers by a
mobile application. This application works completely in offline which use mobile camera to
identify flowers. It will be really helpful for Traveler, Hiker and Nature Enthusiasts [1].

©Daffodil International University 8

2.5 CHALLENGES

Duplicate flower identification is the main challenge of our project. Some flowers are same to look
at by size and shape. That time it is so tough to provide the accurate result. We are continue our
research to solve this problem.

Another challenge is to identify same flower of different specific color. Ex- White Rose, Pink
Rose. This is a tough work to identify color using convolutional neural network. But we take it as
a challenge.

Other challenge is achieving accurate data. In the market of false and duplicate data, it is very
hard to get all of the things right. But we are trying are heart and soul to do better.

©Daffodil International University 9

 CHAPTER 3
RESEARCH METHODOLOGY

AND IMPLEMENTATION

3.1 BASIC INTRODUCTION TO MACHINE LEARNING

In Artificial Intelligence Learning is a very important feature. Many scientists tried to give a proper
definition for learning. Many computer scientists, sociologists, logicians and others discussed
about this for a long time. Some scientists think that learning is an adaptive skill that can perform
the same process better later on (Simon 1987). Others claim that learning is a process of collecting
knowledge (Feigenbaum 1977). Although there is no definite definition for learning skills, we still
have to define machine learning. In general, machine learning has to be identified on how to
improve the computer algorithm automatically through experience (Mitchell 1997).
Machine learning is one of the important field of Artificial Intelligence. At the beginning of
development of Artificial Intelligence (AI), the system does not have a thorough learning ability
so the whole system is not perfect. For instance when the computer faces problems, it can not be
self-adjusting. Moreover, the computer cannot automatically collect and discover new knowledge.
Therefore, computer only can conducted by already existing truths. It does not have the ability to
discover a new logical theory, rules and so on [3].

Environment Learning Knowledge Execution

Figure 3.1: Learning System Structure

©Daffodil International University 10

3.2 CONVOLUTIONAL NEURAL NETWORK

Convolutional neural networks are a class of machine learning networks which are commonly
applied to image visualization problems such as classification. CNNs were inspired by the
connections of the neurons and synapses in the brain. The design of these networks is made up of
series of convolutional, pooling, and fully connected layers. The convolutional layer does what its
name describes, it applies a number of convolutional filters to the input images in order to acquire
the learning parameters for the network. Pooling layers are placed in between convolutional layers,
and are used to reduce the number of parameters used for learning, and thus reduce the computation
required. Finally, fully connected layers are full connections to the previous layer, rather than the
small window the convolutional layers are connected to in the input. Convolutional neural
networks are commonly used for image classification, however, there are limitations to this
application. A human can identify the contents of certain images much more quickly than a
computer, but CNNs have proven to have a 97.6% success rate when applied to facial recognition
[11].

The following figure showing the typical design of convolutional neural network.

Figure 3.2: Typical CNN design

©Daffodil International University 11

3.3 DESIGN OF DATASET

The diagram of dataset design which shows different step of data collection and dataset handler.

Shapla
Shapla
Google Image
Noyontara Noyontara
Downloader
Aporajita
---
Aporajita

CNN Dataset Handler

Shapla Noyontara Aporajita

Figure 3.3: Design of Dataset

©Daffodil International University 12

3.4 DATA COLLECTION PROCEDURE
There are a number of existing datasets which have images of specific flowers. These datasets
were generally collected for very specific uses with neural networks that were designed to classify
flowers based on certain characteristics. The dataset for this project was produced by searching
Google Images using a Python script adapted from a web-crawler created by hardikvasa. As a
result, the images of the flowers are a diverse collection of plants in their natural setting. This adds
the benefit of training the network for use outdoors. The script had to be modified so that it would
receive the list of search keywords from a text file. Google Image Downloader is also helpful to
collect specific flower data on windows platform. The search keywords for the dataset are a long
list of flower observations in Bangladesh. The search keywords use the full genus-species
classification in order to increase the quality of the dataset. For example, rather than searching for
the word “Gardenia”, which would produce undesired images, the preferred keyword would be
“Gardenia jasminoides”. The labels for classification, however, are grouped by genus so that each
class has 400 to over 1000 images.

The following figure shows the sample dataset image of Shapla flower.

Figure 3.4: Dataset of Shapla

©Daffodil International University 13

3.4 STATISTICS OF DATA

The following table contains the statistics of raw data

Flower Name Amount of Training

Image
Shapla 400
Noyon Tara 500
Tulip 800
Aporajita 400
Rose 500
Kadam 450
Daisy 700
Cosmos 350
Apple Flower 300
Rangan 300

Table 3.4 : Statistics of Raw Data

3.5 IMPLEMENTATION REQUIREMENTS (Minimum)

Software

 Windows / Linux / Mac OS

 Python pip package / Anaconda
 Tensorflow
 Android Studio

Hardware

 Intel i3 Processor
 4 GB RAM
 1TB Hard Disk

©Daffodil International University 14

3.6 IMPLEMENTATION OF NETWORK DESIGN

Convolutional Layer 1

Input:

The image data is reduced to a size of 128x128 pixels in order to not overwhelm the hardware
the program was normally tested on. Batches of 32 images are fed into the convolutional layer
and 16 filters of 8x8 pixels are applied to the images.

conv_layer1 =
lb . build_convolutional_layer ( input = image_placeholder ,
num_channels = NUM_CHANNELS ,
filter_size = FILTER_SIZE ,
num_filters = NUM_FILTERS)

Pooling layer 1

Input:

Each pooling layer uses a pool size of 2x2 and a stride size of 2.
pool_layer1 = tf . layers . max_pooling2d ( inputs = conv_layer1 ,
pool_size =[ 2 , 2 ], strides = 2)

Convolutional Layer 2

The second convolutional layer has the same parameters as the first.
conv_layer2 =
lb . build_convolutional_layer ( input = pool_layer1 ,
num_channels = NUM_FILTERS ,
filter_size = FILTER_SIZE ,
num_filters = NUM_FILTERS)
Pooling layer 2

Fully connected layer 1

Each fully connected layer performs an activation on each of its inputs. The first, however,
performs a RELU activation function on the data.

©Daffodil International University 15

connected_layer1 =
lb . create_connected_layer ( input = flat_layer,
num_inputs = flat_layer . get_shape ()[ 1 : 4 ]. num_elements (),
num_outputs = 32 ,
use_relu = True)

Fully connected layer 2

The second FC layer does not perform the RELU activation.

connected_layer2 = \
lb . create_connected_layer ( input = connected_layer1 ,
num_inputs = 32 ,
num_outputs = num_classes ,use_relu = False)

Training Step

The number of training steps can be specified as a command line parameter. Each training step is
validated and tested, and the results of each step are printed to standard out.

is_last_step = (i + 1 == FLAGS.how_many_training_steps)
if (i % FLAGS.eval_step_interval) == 0 or is_last_step:
train_accuracy, cross_entropy_value = sess.run(
[evaluation_step, cross_entropy],
feed_dict={bottleneck_input: train_bottlenecks,
ground_truth_input: train_ground_truth})
tf.logging.info('%s: Step %d: Train accuracy = %.1f%%' %
(datetime.now(), i, train_accuracy * 100))
tf.logging.info('%s: Step %d: Cross entropy = %f' %
(datetime.now(), i, cross_entropy_value))
validation_bottlenecks, validation_ground_truth, _ = (
get_random_cached_bottlenecks(
sess, image_lists, FLAGS.validation_batch_size, 'validation',
FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor,
decoded_image_tensor, resized_image_tensor, bottleneck_tensor,
FLAGS.architecture))
validation_summary, validation_accuracy = sess.run(

©Daffodil International University 16

 [merged, evaluation_step],
feed_dict={bottleneck_input: validation_bottlenecks,
ground_truth_input: validation_ground_truth})
validation_writer.add_summary(validation_summary, i)
tf.logging.info('%s: Step %d: Validation accuracy = %.1f%% (N=%d)' %
(datetime.now(), i, validation_accuracy * 100,
len (validation_bottlenecks)))

©Daffodil International University 17

3.7 DESIGN DEVELOPMENT
The initial step of this project was to research the available machine learning libraries,
convolutional neural network design, and collect datasets. Though there exist some other libraries,
but Tensorflow was chosen because there are many tutorials and documentation for the library.
After began to have a basic understanding of the Tensorflow library, we retrained the MobileNet
model with our own dataset, which proved to be very successful in testing, however, the goal of
this project was to learn how to develop and optimize a neural network. The initial designs of the
CNN for this project were based on several different tutorials about how to use Tensorflow to
design an image classifier. The next step was to begin modifying the initial network to try and find
a design that worked for the application of this project. The next step was implementation of
android app using Android Studio. Some open source code of Tensorflow collected from github
which needed in our implementation work. Final Step was set up the pb file and label file with the
application.

The design of CNN which we tested is being submit as a deliverable for this project. This model
also was trained using the dataset at the genus-species level, which uses approximately 700 images
average per class, which is almost enough. But we are working to add more images per class. This
model is currently being trained and tested. The parameters of the network such as number of
training steps, output directory, and image input directory can all be specified, however, their
defaults will place all the output directory in the current working directory. The input image
directory must be specified, and the contents of the directory must be folders of images in tf files
folder. The Tensorflow trainer will generates two text files: one containing the labels for the
classifier, and the other lists which images were selected for training, testing, and validation which
called pb file. The classifier uses these to read result for each image classification and show the
output result.

©Daffodil International University 18

 CHAPTER 4

EXPERIMENTAL RESULTS AND DISCUSSION

4.1 INTRODUCTION TO IMAGE CLASSIFICATION

The Python script which uses to the trained Tensorflow model is very simple, because most of the
time spent on this project was for collecting the data and learning how to design a CNN. The
classifier is designed to take a directory of images, a text file of the labels used in the network, and
the trained model itself as inputs. The classifier tests the images with the specified model and
displays the results comparing the correct label with the top four classes based on the confidence
level of the predictions [1].

The following figure shows the pre processing step in Tensorflow which create label file and
bottleneck file.

Figure 4.1: Image pre processing by tensorflow

©Daffodil International University 19

The following figure shows the training step which generates the train accuracy value, cross
entropy and validation accuracy value.

Figure 4.2: Training Step

4.2 EXPERIMENT RESULTS

The results of this project is almost successful and has the potential for future improvements. Now
the application can identify around 10 Local Flowers of Bangladesh along with some foreign
flowers. We deeply focused to the accuracy rate. The accuracy rate depends on the amount of data.
So that we use more images with different angel to improve the confidence level. Currently some
flower identify with 100% confidence level. It is one of the success of our research and project.
We used almost 700 images for per flowers for training step. The dataset contains around 5000
flower images. The CNN and the classifier are inconsistent, with some tests resulting in nearly
100% confidence during a correct classification, and other tests which entirely fail to produce a
correct classification. Improvements which can be made are explained in detail later on.

©Daffodil International University 20

The following figures showing the experiment results in real time by using mobile camera.

Figure 4.3: Experiment Result using Android Application

©Daffodil International University 21

4.3 MAIN GRAPH OF TENSORFLOW COMPUTATION

Figure 4.4: Main Graph of Tensorflow Computation

©Daffodil International University 22

4.3 DESCRIPTIVE ANALYSIS

TENSORBOARD

TensorFlow Computation for training a massive deep neural network - can be complex and
confusing. To make it easier to understand, debug, and optimize TensorFlow programs, there
include a visualization tools called TensorBoard. TensorBoard visualizes the quantitative metrics
of the plot, graph, plots and additional information such as passing the images [12].

There are some figures below which shows the histograms of tensorflow computations in
tensorboard.

Figure 4.5: Train Accuracy Graph

©Daffodil International University 23

 Figure 4.6 : Cross Entropy Graph

Figure 4.7: Final Training Summaries Graph

©Daffodil International University 24

The following figures shows the activation histograms and summaries for train and validation
data.

Figure 4.8: Activation Graph

Figure 4.9: Final Training Summaries Histogram

©Daffodil International University 25

 The following figure shows the final training summaries histogram.

Figure 4.10: Final Training Summaries Histogram

©Daffodil International University 26

Figure 4.11 : Train Accuracy Value Figure 4.12: Cross Entropy Value

4.4 EXPERIMENT SUMMARY

Our developed application can identify local Bangladeshi flowers with some foreign flowers also.
We deeply focused on the accuracy level. The accuracy rate depends on the amount of data. So
that we use more images with different angel to improve the confidence level. Currently some
flower identify with 100% confidence level. It is one of the success of our research and project.
We used almost 700 images for per flowers for training step. The dataset contains around 5000
flower images.

©Daffodil International University 27

 CHAPTER 5

5.1 SUMMARY OF THE STUDY

Flower is a very important part of nature. Mostly we identify a plant through its flower.
Experienced botanists do this identification of flower but a naive person will have to consult flower
guidebooks or browse any relevant web pages on the Internet through keywords searching. This is
a system that recognizes the flower in real time using mobile camera. Presently this Android app
can identify around 10 flowers. Most important thing that this app fully works in offline. We are
continuously working to add more flowers to identify. Everyday we see a huge number of flower
species in our house, parks, roadsides, in farms, on our rooftop but we have no knowledge of that
flower species or their origin. Even we have no idea about its name. There are several guidebooks
for flowers knowledge but it becomes quite difficult to find the name when have the picture. Even
the Internet sometimes is not useful. But it is quite difficult for human brain to memorize all the
species they see. Even some flower is similar to look at. This software recognizes the flower in
real time by using mobile camera [5].

The purpose of this project is to use Tensorflow, an open-source dataflow and machine learning
library, to build an image classifying Convolutional Neural Network (CNN) for classifying the
flower image. Tensorflow, in addition to providing developers a simple way to build neural
network layers, can also be run on mobile platforms such as Android. The ultimate goal of our
project is to design and optimize a convolutional neural network for use with flower classification,
and eventually build a simple classification app for mobile devices around the trained network.
The mobile app will allow users to try and classify plants while outdoors or offline [12].

©Daffodil International University 28

5.2 FUTURE WORK

This project has plenty of room for future work, by myself or a future interested student.

1. Improved CNN design. There much more

research and practice is needed to optimize the design.

2. Improve the dataset and add more data.

3. Specific Identification of duplicate flower which is same to look at.

4. We have plan to release this app on Google Playstore.

5.3 CONCLUSIONS
With the rapid development of technology, AI is being used in various fields. Machine learning is
the most basic method to achieve AI. This research describes the work principle of machine
learning and an application of machine learning. At the beginning of development of Artificial
Intelligence (AI), the AI system does not have a thorough learning ability so the whole system is
not perfect. For instance when the computer faces problems, it can not be self-adjusting. Moreover,
the computer cannot automatically collect and discover new knowledge. Therefore, computer only
can conducted by already existing truths. It does not have the ability to discover a new logical
theory, rules and so on. The ultimate goal of this project is to design and optimize a convolutional
neural network for use with flower classification, and eventually build a simple classification app
for mobile devices around the trained network. The mobile app will allow users to try and classify
plants while outdoors or offline. We will continue our research to make the system more efficient.

©Daffodil International University 29

REFERENCES

[1] Kody G. Dangtongdee and Dr. Franz Kurfess, “Plant Identification Using Tensorflow, ”.
Computer Engineering Department, California Polytechnic State University, San Luis Obispo,
Senior Project Final Report. CPE 462, Winter 2018, pp. 5-15.

[2] Xuanxin Liu, Fu Xu, Yu Sun, Haiyan Zhang, and Zhibo Chen, “Convolutional Recurrent
Neural Networks for Observation-Centered Plant Identification,” Journal of Electrical and
Computer Engineering, vol. 2018, pp. 1–7, 2018.

[3] Yu Yang and Patric Granholm, “A study of pattern recognition of Iris flower based on Machine
Learning, ”. Turku University of Applied Sciences, Turku Finland , Bachelor’s Thesis, 2013.

[4] Wang-Su Jeon and Sang-Yong Rhee, “Plant Leaf Recognition Using a Convolution Neural
Network, ”. International Journal of Fuzzy Logic and Intelligent Systems Vol. 17, No. 1, March
2017, pp. 26-34.

[5] Er. Gurleen Kaur Marwa and Er. Salam Din, “Development of Software for Flower Recognition
Using Image Processing Technique, ”. Dept. of SEEIT, COAET, PAU, Ludhiana, India,
vol. 3, Issue 5, May 2015.

[6] Vijaya Bylaiah, “Leaf Recognition and Matching with MATLAB, ”. A Thesis Presented to the
Faculty of San Diego State University, Spring 2014.

[7] Y. Li, Q. Zhu, Y. Cao, and C. Wang, “A leaf vein extraction method based on snakes technique,”
in Proceedings of the International Conference on Neural Networks and Brain (ICNN&B '05), pp.
885–888, 2005.

[8] Bache, K.& Lichman, M. 2013. UCI Machine Learning Repository [http://archive.ics.uci.edu/ml].
Irvine, CA: University of California, School of Information and Computer Science.

[9] Bishop, C. 2006. Pattern Recognition and Machine Learning. New York:Springer, pp.424-428.

[10] Robin. 2010. Machine-Learning - Artificial Intelligence. Available at:

http://intelligence.worldofcomputing.net/category/machine-learning. Consulted 22 Nov 2018.

[11] “Convolutional Neural Network.” Wikipedia, Wikimedia Foundation, Available:

en.wikipedia.org/wiki/Convolutional_neural_network. Accessed on: Nov. 2, 2018.

[12] “TensorFlow.” Wikipedia, Wikimedia Foundation, Available:en.wikipedia.org/wiki/TensorFlow.

Accessed on: Nov. 2, 2018.

©Daffodil International University 30

APPENDICES

Accuracy Value For Some Training Steps.

Number Step Value

1 5 0.46
2 15 0.85
3 18 0.88
4 24 0.91
5 40 0.91
6 42 0.95
7 44 0.95
8 45 0.97
9 47 0.96
10 53 0.96
11 54 0.96
12 56 0.89
13 65 0.97
14 71 0.94
15 72 0.95
16 73 0.99
17 74 0.95
18 79 0.98
19 87 0.96
20 94 0.94
21 102 1
22 105 0.97
23 106 0.99
24 113 0.93
25 118 0.96
26 126 0.95
27 129 0.98
28 130 1

©Daffodil International University 31

 29 139 0.97
30 141 0.98
31 142 0.93
32 150 1
33 155 0.96
34 158 0.97
35 167 0.98
36 168 0.99
37 177 0.97
38 178 0.97
39 183 0.99
40 185 1
41 186 0.99
42 190 0.99
43 198 0.97
44 206 0.98
45 207 0.98
46 208 1
47 212 0.99
48 223 0.98
49 229 0.99
50 234 0.96
51 237 0.98
52 249 0.96
53 255 0.97
54 260 1
55 263 0.98
56 267 0.99
57 271 0.98
58 273 0.96
59 292 1
60 297 1

©Daffodil International University 32

 61 298 0.99
62 303 1
63 307 0.99
64 308 1
65 310 0.98
66 318 0.99
67 319 0.98
68 325 1
69 326 0.99
70 327 1
71 330 1
72 339 1
73 347 1
74 363 0.99
75 367 0.99
76 373 0.99
77 377 0.99
78 379 0.99
79 391 1
80 392 1
81 408 1
82 421 0.99
83 422 1
84 424 0.99
85 430 1
86 439 1
87 441 0.98
88 442 0.99
89 445 1
90 451 0.97

©Daffodil International University 33

View publication stats