International Journal of Recent Technology and Engineering (IJRTE)

ISSN: 2277-3878 (Online), Volume-8 Issue-6, March 2020

Tensorflow Based Image Classification using

Advanced Convolutional Neural Network
Pradumn Kumar, Upasana Dugal

Abstract: In this Research study image identifications will be

done by the help of Advanced CNN (Convolutional Neural Quantitative features provide counted or measured data while
Networks with Tensorflow Framework. Here we use Python as a Qualitative features gives in terms of their properties like
main programming language because Tensorflow is a python shape, colour, texture etc. All these things require
library. In this study input data mainly focuses on Plants classification because more than one plant look exactly same
categories by the help of leaves for identifications. Selecting CNN to each other [11].
is the best approach for the training and testing data because it Taxonomists want more efficient methods to fulfill the
produces promising and continuously improving results on requirements for identification [9] because the image
automated plant identifications. Here results are divided in terms identification is biggest challenge in Machine visions. So in
of accuracy and time. Using advanced CNN results are above this paper Tensorflow based CNN is used for identification,
95% while on others accuracy is below 90% and taking much here thousands of images should be use in the form of
time than this.
datasets.
Keywords: Tensorflow, CNN, Image Identification.
II. RELATED WORK
I. INTRODUCTION
This study proposed a new algorithm for plant identification
Currently image identification is fastest developing using fractal dimensions [2]. Here 3- stages of fractal
technology for developers in the form of data. Let us take an requirement had performed of leaves. In these stages
example to better understand that what image identification Contour Nervure and Nervure fractal dimensions are used.
is. Today Google uses image captcha for validation of users. Here the accuracy rate is 84% but have drawbacks. It is fully
Now a days in social media in which there are multiple dependent on clarity of images.
images of users in the form tagged images or untagged In studied about deep learning as a method for image
images. So in social media this technology plays an vital role classification with hand crafted feature approach. In both
to identify users by their facts with 95% and above accuracy approach only valid bias introduced. In this picture of
[1].
speciman is not training it wins realistic to assume. Due to
Now a days technology is beating human’s ability for image
unavailability plant image dataset the accuracy is low and
identifications. Here machine learning having power and
influence over other approach. Machine learning is a part of time consuming [3].
Artificial Intelligence which can perform activities without According to research paper Neural Network Architecture as
human intervention. a method for image classification. In Neural Network
The term biodiversity is very remarkable for obvious features Architecture framework, the combination between mimics
of organic things. In spite of the different categories of of two pairs of human eye and variations sequence auto
organisms can be grouped into the taxonomy. In biological encoding. In this complex images improve the MNIST
terms taxonomy consists with the classes by their names and models. MNIST datasets also used for fashions, automobiles
its behavior and properties, so here one question arises why and so many other image identification.
we are using the term identification. The identification shows This study is about image classification by using deep
the assignment of unknown organic things [14]. learning via framework tensorflow. This classification is
In this study we focus on plants identification, it is the performed on 5- different types of flowers with an accuracy
process of assigning a particular plant to taxonomy according of up to 90% but in this there is drawback that if the size of
to their characteristics. These characteristics will be smaller then accuracy was slightly low as compared to
identified by the help of Quantitative and Qualitative bigger.
features. This research paper introduced plant leaf species
identification using Curvelet transform with Support Vector
Machine. In this digital image of leaves divided into 25 sub
images and the SVM is used for classification. On
comparing with other method, it was good relatively slow
Manuscript received on February 10, 2020. and less efficient [4].
Revised Manuscript received on February 20, 2020. In this study Multilevel Deep learning architecture is used
Manuscript published on March 30, 2020. for classification of land cover and crop types using remote
* Correspondence Author
Pradumn Kumar, Department of Computer Science & sensing data. Here optical image classification is done by
Engineering,Babu Banarasi Das University, Lucknow, India. Email: the unsupervised learning and to restore the missing data
[email protected] and supervised entities. As a results 2-D CNNs depicts
Upasana Dugal, Department of Computer Science & Engineering, highest accuracy for certain supervised data. The target
Babu Banarasi Das University, Lucknow, India. Email:
[email protected]. accuracy was 85% but in actual results it was 94.6%. In this
study if the object is small in size then image misclassifies
© The Authors. Published by Blue Eyes Intelligence Engineering and [5] [12].
Sciences Publication (BEIESP). This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Retrieval Number: F7543038620/2020©BEIESP

DOI:10.35940/ijrte.F7543.038620 Published By:
Journal Website: www.ijrte.org Blue Eyes Intelligence Engineering
994 & Sciences Publication
 Tensorflow Based Image Classification using Advanced Convolutional Neural Network

III. CHALLENGES IN IMAGE-BASED TAXONOMY There are approximately between 750 – 780 operations with
IDENTIFICATION different modules independently in whole training graph.
There are generally three steps in training graphs:-
A. Huge Number of taxonomy to be Biased Model Inputs: Read operations and preprocess CIFAR
In this world there are large number of species available like images operations will be added for evaluation and training
plants, animals, insects and so many others. It is very respectively.
difficult to prejudice them into taxonomical form. In any
Model Prediction:- On supplied images classifications
event, while confining the concentration to the verdure of an
should be done by adding operations that perform inferences.
region, thousands of flora and fauna categories should be
upheld. Let us take an example of German’s Flora exhibits Model Training:- Add operations that compute the loss,
thousands of flowering species and in each single genus there gradients, variable updates and visualization summaries.
will be many other species.
B. Variations on Huge Intraspecific A. Training a Model Using Multiple GPU Cards
Some species plants have some different horticulture In scientific operations of computer multiple GPUs are used
characteristics like location, moisture, nutrition, life cycle for modern workstations. Tensorflow can influence this
and climatic conditions etc. These changes in horticulture surroundings to run the training operation simultaneously
characteristics can occur on their measuring units , flowers, across multiple GPUs cards. A proper training processes are
fruits, leaves and sometimes even whole plants. For example required to run a training model in a parallel or distributed
Knautia arvensis commonly known species of knautia genus. manner.
This ranges from huge whole or dentate lanceolate ground Here we entitled model replication to be one replica of a
leafs over profoundly lobed and practically pinnate stem model training on a subset of data. Naively the usage of
leafs to little and again lanceolate and whole upper stem leafs Asynchronous updates of model parameters results in sub-
greatest training overall performance due to the fact an
C. Accession process for variation individual model replication is probably be trained on musty
Plants leaves are in 3-d images in reality but when we replica of the model parameters. Conversely using
capture then it becomes in 2-d so sine of the features change completely synchronous updates might be as gradual because
like appearance. By this large differences create between the slowest version replica.
original images and captured images like shape and Using multiple GPUs in modern workstation, each GPU
appearance. External conditions of image capturing also contains same processing speed and enough memory to run
limited like zoom, focus, sensor, and resolutions etc. CIFAR models. Therefore we will design our training model
as follows:-
D. Variation on small intraspecific  Import an individual model replication on every
In the species look similar to each other but there is GPU.
challenging task to distinguish them. These variations are
 Updating should be done in synchronously
invisible characteristics so in these situation horticulture
manner which means wait for all GPUs to finish
characteristics are very crucial to prejudice. Even sometimes
their batch data.
experts are challenged to distinguish the species in safely
manners by almost invisible characteristics. For example,
accurate distinguishing between species required visible B. Proposed Method
characteristics like flower or fruits but all these are According to this architecture image classification will be
seasonable things. So out of season its very difficult to done by CNN. In this whole framework there are 7 stages
distinguish and each stages has their own discussion. All these stages
depend on Tensor flow function which is an open source
IV. METHODOLOGY software and all the Tensorflow libraries are on python
programming language and on importing Tensorflow each
Specially using CNN is a very trending procedure for Deep stage process will be done as per design [6] [7][15].
learning in computer point of view. ImageNet have produced
a lot of expectation by giving exciting results [9]. Here CNN
takes the most challenging task for identification of plants by
using their complete picture or any parts of that plants while
others tackles one by one process like firstly they take any
specific organisms (flowers, leaves and bark etc.) then whole
picture of organisms. In CNN there are some limitations like
it is not better with very large sets of images or lack of
explanatory power.
So Advanced CNN will replace CNN because in Advanced
CNN is small in size as compare to CNN for recognizing
images. Here large models can be easily scale up and these
models are small enough to train fast, by this we will get out
new ideas and have a good chance for experiment on other
methods also.
The architecture of Advanced CNN is multi-layer consisting
of alternate use of Convolution layers and nonlinearities. All
these layers are followed by fully connected layers leading
into a softmax classifier. This model gives a good accuracy
results with in few time when we run on a GPU.
Retrieval Number: F7543038620/2020©BEIESP
DOI:10.35940/ijrte.F7543.038620 Published By:
Journal Website: www.ijrte.org Blue Eyes Intelligence Engineering
995 & Sciences Publication
 International Journal of Recent Technology and Engineering (IJRTE)
ISSN: 2277-3878 (Online), Volume-8 Issue-6, March 2020

To perform classification we will use one or more

dense layer on the output tensor of the convolutional base.
The output will be in 3-D form and dense layers will take
into vector form i.e. 1-D. so firstly we have to unroll or
flatten 3-D output into 1-D then addition of one or more
dense layer on the tip. As we know CIFAR has 10 output
classes so we use final dense layer which 10 outputs and a
SoftMax activation layer [13].
5. Compile and Train the Model:-
Feeding the trained data to the mode which
means two arrays will be create train images and train labels.
Association of images and labels into the model. Predictions
will be performed about a testsets.
To start training an epoch in Machine
learning is the complete processing via the studying set of
rules of the complete train-set. In epoch these is one training
iterations and one iteration will iterate all the sample once
after calling tensorflow.
Train function and outline the value for the
parameter epochs, you determine how generally model need
to be trained on sample data(commonly at least some
hundred instances).
6. Evaluation of Model:-
1. Download and Prepare CIFAR-10 Dataset :-
In CIFAR-10 dataset there are 60000 images which are in 10 Here Image classifier will train towards better
classes and in each class there will be 6000 images. Here the accuracy with less time.
dataset will be divided into two phase first is training and
second is testing phase.In training phase there are 50000
images while in testing phase there are 10000 images. Here 1
all the classes are mutually exclusive and no overlapping
0.9
between them.
0.8
2. Verification of Data :-
0.7
Verification of dataset plays a vital role here. This
verification will be done to check whether the given dataset 0.6
is correct or not. For verification few images will used for Accuracy 0.5
plotting and these images are from training phase and each Accuracy
0.4
image will be shown by their specific class name. Epoch
0.3
3. Create the Convolutional Base :-
0.2
Here Convolutional Base follows common pattern
in form if Con2D and maxPooling2D layers in a stack. In this 0.1
CNN takes input in the form of tensor shape (RGB) which 0
means (image_height, image_width, colour_channels) [8]. 2 4 6 8
As per above the output of Conv2D and maxPooling2D Epoch
layers are in 3-D tensor shape. If you go deeper in the
Network the height and width will be shrinked. The output of
each Conv2D controlled by the arguments, if height and
width is shrinked then we have to add more output channels
in every Conv2D layers. C. Image Classification Flowchart:-

4. Adding Dense Layer on Top :-

Retrieval Number: F7543038620/2020©BEIESP

DOI:10.35940/ijrte.F7543.038620 Published By:
Journal Website: www.ijrte.org Blue Eyes Intelligence Engineering
996 & Sciences Publication
 Tensorflow Based Image Classification using Advanced Convolutional Neural Network

identification, recognition etc. Here results are very

optimized as compare to DNN. But in CNN validation
loss is high which causes over-fitting.
3. Transfer Learning is another approach which are used
for reusing the acquired knowledge. It means already
trained model are used on large dataset to get well results
on related works. But here accuracy is good and less
time comparison to others.
But further we can improve our accuracy and time
management by adding more data augmentation, more
epochs and most importantly adding layers. So Advanced
CNN is complete replacement of all these.

TABLE I. RESULTS FOR IMAGE

CLASSIFICATION MODELS
Accuracy Time Error Validation
rate Consume rate loss
DNN 70-80% 6.4Hrs Very 7.8
high
CNN 90% 5.4Hrs High 3.3

Transfer 92% 12mins Low 0.64

Learning
Advanced More than 8mins Very 0.3
CNN 95% low

VI. CONCLUSION AND FUTURE DIRECTION

In conclusion, this research study we have discussed about
image identification or classification by using Advanced
CNN via Tensorflow Framework. In this study we have
performed classifications on leaves of plants by using
CIFAR 10 dataset. As results we check the comparison
between multiple models with specified dataset. All the
results are achieved as per objectives by the Advanced CNN
The implementation of image classification flowchart will be with accuracy of more than 95% while others are not
done using Tensorflow. Here flowchart depicts that the capable to give results as per objective. Advanced CNN is
classification will start from examining and understand the our main agenda for image classification because in this
data. After that input pipeline will be built then CNN is adding dense layers and increasing epochs gives desired
applied to train the model. In CNN testing depends on the results in better way. Epochs are used to control the over
images of leaves and if the output is not according to your fitting problems. Advanced CNN are very faster in
expected results then it needs to restart the CNN to get comparison to other, it take very less time for classification.
accurate results. This process will end when output is Because Advanced CNN works on GPUs and will also work
classified into their specified category. their own TPUs. TPU is even more faster than the GPU. So
by this we will get better results than others. We will further
V. COMPARISON OF IMAGE CLASSIFICATION improve our Advanced CNN for classification for large
MODELS & RESULTS number of images and even we can modify our model.
Tensorflow framework is wide technology to create data
In Deep learning models there are various image models so research will continues on this by providing large
classifications models which are used in practical number of images of species.
applications. Many methodologies have been created and
even still popping up. So here we will introduce some
basics of other models in comparison with our Advanced
CNN.
1. Deep Neural Networks(DNN) is used to train a neural
networks for regression and classification. The
performance of DNN is not well with the images
because the accuracy is bad.

2. Convolutional Neural networks (CNN) represented to be

very successful in image classification, object

Retrieval Number: F7543038620/2020©BEIESP

DOI:10.35940/ijrte.F7543.038620 Published By:
Journal Website: www.ijrte.org Blue Eyes Intelligence Engineering
997 & Sciences Publication
 International Journal of Recent Technology and Engineering (IJRTE)
ISSN: 2277-3878 (Online), Volume-8 Issue-6, March 2020

REFERENCES
Upasana Dugal received Master’s degree from
1. Mohd Azlan Abu1, Nurul Hazirah Indra1, Abdul Halim Abd Rahman1, Karnataka, India. Currently she serves as Associate
Nor Amalia Sapiee1 and Izanoordina Ahmad1, “A study on Image Professor at the Department of Computer Science at
Classification based on Deep Learning and Tensorflow”. International Babu Banarasi Das University of Lucknow, India.
Journal of Engineering Research and Technology. ISSN 0974-3154, She has 14 years’ experience of teaching. Her area of
Volume 12, Number 4 (2019), pp. 563-569 interest in research includes Machine Learning,
2. Antony Jobina,*, Madhu S. Nairb, Rao Tatavarti, Plant Identification Natural Processing language, Design and Analysis of
Based on Fractal Refinement Technique(FRT). 2nd International Algorithms, Theory of Automata and Formal Languages and Computer
Conference on Communication, Computing & Security [ICCCS-2012] Graphics. She has participated in many research projects for multiple
3. Trishen Munisami, Mahess Ramsurn, “Plant leaf recognition using journals and she holds Various certifications from IBM and Indian Institute
shape features and colour histogram with k-nearest neighbor of Technology(IITs).
classifiers”. Second International Symposium on Computer vision and
the Internet(VisionNet”15).
4. Shitala Prasad, Piyush Kumar, R. C. Tripathi, “PlantLeaf Species
Identification using Curvelet Transform”. International Conference on
Computer & Communication Technology (ICCCT)-2011.
5. Nataliia Kussul, Mykola Lavreniuk, Sergii Skakun, and Andrii
Shelestov, “Deep Learning Classification of Land Cover and Crop
Types Using Remote Sensing Data” IEEE GEOSCIENCE AND
REMOTE SENSING LETTERS, VOL. 14, NO. 5, MAY 2017.
6. Heba F. Eid1,∗ and Aboul Ella Hassanien2,3,∗ and Tai-Hoon Kim4,
“Leaf plant identification system based on hidden naive bays classifier”
2015 4th International Conference on Advanced Information
Technology and Sensor Application.
7. Kody G. Dangtongdee,Dr. Franz Kurfess, “Plant Identifica-tionUsing
Tensorflow”. Plant Phenomics Volume 2019, Article ID 9237136, 14
pages https://doi.org/10.34133/2019/9237136CPE 462, Winter 2018
8. Marco Seeland1* , Michael Rzanny2, David Boho1, Jana Wäldchen2
and Patrick Mäder1,”Image-based classification of plant genus and
family for trained and untrained plant species”. Institute for Computer
and Systems Engineering, Technische University Ilmenau,
Helmholtzplatz 5, 98693 Ilmenau, Germany. 2Max-Planck-Institute for
Biogeochemistry, Department Biogeochemical Integration, Hans-
Knöll-Str. 10, 07745 Jena, Germany.
9. Jana Wäldchen , Michael Rzanny, Marco Seeland, “Automated plant
species identification—Trends and future directions”. PLoS Comput
Biol 14(4): e1005993. journal.pcbi.1005993 & April 2018 .
10. Faith Ertam, Galip Aydin,”Data classification with deep learning using
Tensorflow”. 2nd International Conference On computer Science and
Engineering.ISSN 978-1-5386-0930-9/17.
11. Jonathan Y.Clark, David P.A. Corney and H.LIlian Tang,”Automated
Plant Identification Using Artificial Neural Networks”.presented at
CIBCB2012 @IEEE12.
12. Atharva Sharma a,*, Xiuwen Liu a, Xiaojun Yang b, Di Shi c, “A
patch-based convolutional neural network for remote sensing image
classification” Neural Networks 95 (2017) 19–28.
13. Kanit Wongsuphasawat en al, “Visualizing Dataflow Graphs of Deep
Learning Models in TensorFlow” IEEE Transactions On Visualization
And Computer Graphics, Vol. 24, No. 1, January 2018
14. Xiaoling Xia, Cui Xu, Bing Nan,” Inception-v3 for Flower
Classification” 2017 2nd International Conference on Image, Vision
and Computing, 978-1-5090-6238-6/17
15. Zixian Zeng1 ,Qingge Gong2 ,Jun Zhang1,” CNN Model Design of
Gesture Recognition Based on Tensorflow Framework”2019 IEEE 3rd
information Technology,Networking,Electronic and Automation
Control Conference (ITNEC 2019).

AUTHORS PROFILE

Pradumn kumar holds a B.Tech in Computer Science

& Engineering from Greater Noida, India. He is
currently pursuing M.Tech in Software Engineering at
Babu Banarasi Das University of Lucknow in India.
His research interests include machine learning with
Tensorflow framework for image classifications. He
has knowledge of various programming skills which
requires for his area of interests. He is a member of
Institution of Engineers (India).

Retrieval Number: F7543038620/2020©BEIESP

DOI:10.35940/ijrte.F7543.038620 Published By:
Journal Website: www.ijrte.org Blue Eyes Intelligence Engineering
998 & Sciences Publication