International Journal of Recent Technology and Engineering (IJRTE)

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

Image Colour Prediction using Deep learning

K. Vishnu Prakash, S. Siva Prakash, H. Vishnu Harichandran, S Petchiappan, I Muthu Selvi

It's easy to change over a shaded picture with RGB

Abstract: Any image we perceive through a screen is made Channels into a Grayscale picture yet the other way around
of three separate channels, R, G, and B. With the help of these
is generally testing. Anticipating the shade of a picture just
three channels; an image comes to colour. Any pictures taken
during the old times were in grayscale format. To convert any by utilizing the accessible grayscale channel is very
given grayscale image into colour, we need the help of a troublesome as this necessary human ability to recolour the
photoshop professional, which might take hours of the picture by photoshop control. Nonetheless, on the off chance
workforce. In a revolution to this, we propose an utterly that we utilized the broad field of profound figuring out how
programmed methodology that produces lively and practical
to further our potential benefit, it is conceivable to recolour
colourizations. Generative adversarial networks are an
unsupervised learning task in machine learning that involves a dark &white picture to Shading via preparing.
automatically discovering and learning the regularities or
patterns in input file to get an output. In our case, a grayscale II. EXISTING METHODOLOGY
image can be converted to colour with the help of GANs.
In the current scenario if someone has to convert a
Keywords Channels, Colorizations, GAN, Grayscale, black & white image into colour one might require a person
Learning, Photoshop, Unsupervised. with photoshop expertise and hours of hard work. Otherwise
they might need supervised learning algorithms and tons of
I. INTRODUCTION training to deal with vast amounts of data to convert a
simple black & white image into colour and that too with
Deep learning has advanced a great deal in this cutting-edge very little accuracy.
time, which has achieved a blast of information in all Existing system supports colour extraction feature
structures and from each district of the world. Big from a coloured image or it can fill in the shadows and
information is gotten from sources like social media, search highlights with one particular colour, but it is practically
motors, web-based business stages, and online films. This impossible to dynamically colour adapt to any given image.
immense measure of information is effectively available and The drawback with existing systems is that it takes lots of
can be shared through fine tech applications like distributed human power and technical knowledge in order to predict
computing. Be that as it may, the information, which the colour of a grayscale image. Even with regular machine
ordinarily is unstructured, is huge to such an extent that it learning the task of teaching a model is a rigorous task
could take a very long time for people to appreciate it and especially with supervised learning. Supervised learning is
concentrate significant data. Organizations understand the the machine learning task of learning a function that maps an
extraordinary potential that can come about because of input to an output based on example input-output pairs [2]. It
disentangling this abundance of data and are progressively infers a function from labelled training data consisting of a
adjusting to artificial intelligence frameworks for set of training examples.
computerized support.

Manuscript received on February 10, 2020.

Revised Manuscript received on February 20, 2020.
Manuscript published on March 30, 2020.
* Correspondence Author
K. Vishnu Prakash*, UG Student, Department of CSE, National
Engineering, Kovilpatti, India. Email: vishnuprakash71298@gmail.com Fig. 1. Image Colourization using photoshop
S. Siva Prakash, UG Student, Department of CSE, National Engineering, In this fig. 1 it shows how the image has been colourizing
Kovilpatti, India. Email: sivaprakash354@gmail.com using photoshop, It take lot of time and concentrate deeply.
H.Vishnu Harichandran, UG Student, Department of CSE, National
Engineering, Kovilpatti, India. Email: 172402@nec.edu.in
S Petchiappan, UG Student, Department of CSE, National Engineering, III. PROPOSED METHODOLOGY
Kovilpatti, India. Email: 1712402@nec.edu.in
Mrs. I Muthu selvi, Assistant Professor, Department of CSE, National
Image colourization is an image-to-image translation
Engineering, Kovilpatti, India. Email muthu2012skc@gmail.com problem that maps a high dimensional input to a high
dimensional output[1].
© The Authors. Published by Blue Eyes Intelligence Engineering and
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: E5935018520/2020©BEIESP

DOI:10.35940/ijrte.E5935.038620 Published By:
Journal Website: www.ijrte.org Blue Eyes Intelligence Engineering
2037 & Sciences Publication
 Image Colour Prediction using Deep learning

It can be seen as pixel-wise regression problem where Supervised learning algorithms figure out how to outline
structure in the input is highly aligned with structure in the work y'=f(x), given marked information y. A case of this
output. That means the network needs not only to generate would be grouping, where one could utilize client buy
an output with the same spatial dimension as the input, but information (x) and the client's particular age (y) to
also to provide colour information to each pixel in the characterize new clients. The vast majority of the
grayscale input image. We provide an entirely convolutional administered learning calculations are inalienably
model architecture using a regression loss as our baseline discriminative, which implies they figure out how to show
and then extend the idea to adversarial nets[3]. In this work the contingent likelihood dissemination work (p.d.f).
we utilize the LAB colour space for the colourization task. Regardless of the way that one could make expectations
with this p.d.f, one isn't permitted to test new from the
information dispersion legitimately.

Fig. 2. RGB and LAB Colour model

In this fig 2 LAB colour space contains dedicated
channel to depict the brightness of the image and the colour
information is fully encoded in the remaining two channels.
Fig no 3.2 represents LAB colour space. As a result, this
prevents any sudden variations in both colour and brightness Fig. 4. GAN Working model
through small perturbations in intensity values that are In this fig 4 shows the composed two models of Generative
experienced through RGB. Adversarial Network. The primary model is known as a
Generator and it plans to produce new information like the
IV. SYSTEM REQUIREMENTS AND DESIGN normal one. The Generator could be assimilated to a human
We use the jupyter notebook to store our data and its one the craftsmanship falsifier, which makes counterfeit show-
python platform. Colaboratory is a free Jupyter scratchpad stoppers. The subsequent model is named the Discriminator.
condition that requires no arrangement and runs completely This present's model will probably perceive if info
information is 'genuine' — has a place with the first dataset
in the cloud. With Colaboratory you can compose and
— or in the event that it is 'phony' — created by a falsifier.
execute code, spare and offer investigations, and access
Right now, Discriminator is comparable to the police (or a
amazing processing assets, just for nothing from your workmanship master), which attempts to identify
browser. It is unbelievably simple as it makes it share craftsmanship’s as honest or extortion.
journals and model code. One can open another person's
scratchpad and run it immediately, with no of the typical B. Architecture Diagram
obstacles around ensuring you have the correct bundles The architecture of generator is motivated by U-Net[5]: The
installed. This is on the grounds that Collab cases design of the model is symmetric, with n encoding units and
accompany such huge numbers of information science and in interpreting units. Fig. 6 shows the contracting way
AI libraries pre-installed. Fig no 4.1 shows Google comprises of 4x4 convolution layers with walk 2 for down
Collaboratory. sampling, each followed by cluster standardization and
Leaky-ReLU initiation work with the incline of 0.2. The
A. GAN quantity of channels is multiplied after each progression.
GANs have just gotten broadly known for their Every unit in the sweeping way comprises of a 4x4
application flexibility and their remarkable outcomes in transposed convolutional layer with walk 2 for up sampling,
creating data. They have been utilized, all things link with the initiation guide of the reflecting layer in the
considered, applications for content/picture/video age, contracting way, trailed by clump standardization and ReLU
medicate disclosure and content to-picture synthesis[4]. Fig actuation work. The last layer of the system is a 1x1
4.3 speaks to the abilities of the GAN model. convolution which is identical to cross-channel parametric
pooling layer.

Fig. 3. GAN model

Retrieval Number: E5935018520/2020©BEIESP

DOI:10.35940/ijrte.E5935.038620 Published By:
Journal Website: www.ijrte.org Blue Eyes Intelligence Engineering
2038 & Sciences Publication
 International Journal of Recent Technology and Engineering (IJRTE)
ISSN: 2277-3878 (Online), Volume-8 Issue-6, March 2020

We have started with the ImageNet dataset and changed

generally speaking pictures from the RGB concealing spaces
to the lab concealing space[8]. Like the RGB concealing
spaces, the lab concealing space has three channels.
Nevertheless, rather than the RGB concealing region, lab
encodes concealing data in a startling way
 The L channel encodes softness force
 The A channel encodes green and red
 And the B channel encodes blue and yellow

Fig. 5. Image processing format

C. Data Flow Diagram A. Training Gan
Before plunging into the UNET model, it is critical to
comprehend the various tasks that are commonly utilized in
a Convolutional Network fig. 6.

Fig. 8. Training model

Since both the generator and discriminator are being
modeled with neural, networks, a gradient-based
optimization algorithm can be used to train the GAN. The
fundamental steps to train a GAN can be described as
following[9]:

 Sample a noise set and a real-data set, each with

size m.
 Train the Discriminator on this data.
Fig. 6. Data flow diagram  Sample a different noise subset with size m.
 Train the Generator on this data.
V. IMPLEMENTATION AND RESULTS  Repeat from Step 1.

Fig. 9. Final result

After the block and white image was inserted then the GAN
was change to colorized image using Generator and
Discriminator. Generator was used to find the color of the
Fig. 7. Tools used to colourize the image particular place
This application is related to picture taking care of subject
to CNN (Convolution Neural Framework) and significant VI. CONCLUSION
learning[6]. The fundamental idea behind this undertaking is
to change over high complexity pictures to the shaded While picture colorization is a complex task, it is
picture. We are using the Convolution Neural Framework fit nevertheless an occasion of a troublesome pixel forecast.
for concealing profoundly differentiating pictures. Picture Here we have demonstrated that colorization with a
colorization is the path toward taking a data grayscale (high profound CNN and a result that can come nearer to creating
differentiation) picture and a short time later conveying a results indistinct from
yield colorized picture that addresses the semantic genuine photographs[10].
concealing tones of the information[7].

Retrieval Number: E5935018520/2020©BEIESP

DOI:10.35940/ijrte.E5935.038620 Published By:
Journal Website: www.ijrte.org Blue Eyes Intelligence Engineering
2039 & Sciences Publication
 Image Colour Prediction using Deep learning

Our system does not just give a yield. It can even Nadu, India in the year 2017.
be an affectionate task for outline learning and future
I. Muthu Selvi is an Assistant Professor in the
research. Albeit just prepared to shade, our method learns a Department of Computer Science and Engineering at
portrayal that is shockingly valuable for object grouping, National Engineering College, Tamil Nadu, India. She
identification, and division, performing contrasted with received her B.E. Computer Science and Engineering
degree from National College of Engineering, Tamil
other unequivocally self-managed pre-preparing
Nadu in the year 2005. She completed her M.E.
techniques[11]. Computer Science from Anna University of Technology,
Tamil Nadu in 2010. She has 10 years of teaching experience. She is a
REFERENCES active member in IEEE. She has organized many National Level Seminars
and Workshops sponsored by various funding agencies.
1. Agrawal, P., Carreira, J., Malik, J.: Learning to see by moving. In:
Proceedings of the IEEE International Conference on Computer
Vision. (2015) 37–45
2. Bengio, Y., Courville, A., Vincent, P.: Representation learning: A
review and new perspectives. IEEE transactions on pattern
analysis and machine intelligence 35(8) (2013) 1798–1828
3. Charpiat, G., Hofmann, M., Scholkopf, B.: Automatic image
colorization via multimodal predictions. In: Computer Vision–
ECCV 2008. Springer (2008) 126–139
4. Cheng, Z., Yang, Q., Sheng, B.: Deep colorization. In: Proceedings
of the IEEE International Conference on Computer Vision. (2015)
415–423
5. Ngiam, J., Khosla, A., Kim, M., Nam, J., Lee, H., Ng, A.Y.:
Multimodal deep learning. In: Proceedings of the 28th
international conference on machine learning (ICML-11). (2011)
689–696
6. Ramanarayanan, G., Ferwerda, J., Walter, B., Bala, K.: Visual
equivalence: towards a new standard for image fidelity. ACM
Transactions on Graphics (TOG)26(3) (2007) 76
7. E. Reinhard, M. Ashikhmin, B. Gooch, and P. Shirley,
“Colortransfer between images,” IEEE Computer Graphics and
Applications, September/October 2001, pp. 34-41
8. Simonyan, K., Zisserman, A.: Very deep convolutional networks
for largescale image recognition. arXiv preprint arXiv:1409.1556
(2014)
9. T. Welsh, M. Ashikhmin and K. Mueller, “Transferring color to
greyscale images,” in The 29th Annual Conference on Computer
Graphics and interactive Techniques, Texas, 2002, pp. 277 – 280
10. J.Yoo and S.Oh, “A coloring method of gray-level image using
neural network,” in The International Conference on Neural
Information Processing and Intelligent Information Systems, vol.
2, Singapore, 1997, pp. 1203 - 1206.
11. Dahl, R.: Automatic colorization. In http://tinyclouds.org/colorize/.
(2016)

AUTHORS PROFILE
K. Vishnu Prakash is a present final year
Engineering Graduate in the Department of Computer
Science and Engineering at National Engineering
College, Tamil Nadu, India. He did his Higher
Secondary from Viveka Matriculation Higher
Secondary School, Sivagiri, Tamil Nadu, India in the
year 2016.

S. Siva Prakash is a present final year Engineering

Graduate in the Department of Computer Science and
Engineering at National Engineering College, Tamil
Nadu, India. He did his Higher Secondary from Sri
jayendra swamigal silver Jubilee Matriculation Higher
Secondary School, maharaja nagar, Tirunelveli, Tamil
Nadu, India in the year 2016.

H. Vishnu Harichandran is a present final year

Engineering Graduate in the Department of Computer
Science and Engineering at National Engineering
College, Tamil Nadu, India. He did his Diploma from
Jayraj Annapakicam csi polytechnic Margoschis
Nagar,Nazareth, Tuticorin, Tamil Nadu with the stream
of computer science in the year 2017.

S. Petchiappan is a present final year Engineering

Graduate in the Department of Computer Science and
Engineering at National Engineering College, Tamil
Nadu, India. He did his Higher Secondary from Thiru
irutha Higher Secondary School, , Tamil Nadu, India in
the year 2015. He Completed his Diploma from sankar
polytechnic college, sankarnagar, Tirunelveli, Tamil

Retrieval Number: E5935018520/2020©BEIESP

DOI:10.35940/ijrte.E5935.038620 Published By:
Journal Website: www.ijrte.org Blue Eyes Intelligence Engineering
2040 & Sciences Publication