Documente Academic
Documente Profesional
Documente Cultură
BACHELOR OF TECHNOLOGY
In
BY
Submitted to
CERTIFICATE
This is to certify that A Major Project entitled “IMAGE SUPER RESOLUTION USING
DCNN” is a record of bonafide work carried out by the student B.Sanjay Kumar
(15K41A05G1) during the academic year of 2015-2019 in the partial fulfilment for the award of
ACKNOWLEDGEMENT
I wish to take this opportunity to express my sincere gratitude and deep sense of
Respect to our beloved Principal, Dr.V.MAHESH, for his continuous support and guidance
to carry out this project.
I express my heartfelt thanks to our Head of the Department of Computer Science and
Engineering, Mr A.Srinivas,Asst.Professor, for his encouragement, guidance from time -to-
time, and we are thankful to all Faculty members & Programmers of CSE Department of SR
Engineering College for their help during our course.
We would like to express our sincere gratitude to our internal guide Mr D.Ramesh
Sr.Asst.Professor of Department of Computer Science and Engineering, for providing me with
necessary infrastructure and encouragement in everything in our work on this project whose
guidance, and suggestions has motivated me to achieve goals we never thought possible. The
time we have spent working under is truly been a pleasure.
Finally I would like to thank all my B.Tech. Classmates for their constant support during
our class work. And special thanks to our parents and family members for their support and
encouragement throughout completion of the project.
P.Sriya (15K41A05G1)
ABSTRACT
We propose a deep learning method for single image super-resolution (SR). Our method directly
learns an end-to-end mapping between the low/high-resolution images. The mapping is represented as a
deep convolutional neural network (CNN) that takes the low-resolution image as the input and outputs the
high resolution one. Our deep CNN has a lightweight structure, yet demonstrates state-of-the-art
restoration quality, and achieves fast speed for practical on-line usage. We explore different network
structures and parameter settings to achieve trade-offs between performance and speed. Moreover, we
extend our network to cope with three color channels simultaneously, and show better overall
reconstruction quality.
INDEX
1. INTRODUCTION
2. LITERATURE SURVEY
3. DESIGN
4. IMPLEMENTATION
4.1 MODULES
5. TESTING
6. RESULT
7. CONCLUSION
8. FUTURE SCOPE
9. BIBLIOGRAPHY
LIST OF FIGURES
1.
2.
3.
4.
LIST OF TABLES
1.
2.
3.
4.
5.
1. INTRODUCTION
1.1 ABOUT PROJECT
An Image consist of pixels (smallest element of image). Pixels are further stored in memory in
the form of raster image or raster map, which is in 2d array of small integers. When a low resolution
image is up-scaled, the image gets distorted which is very unpleasant. So in order to tackle this problem
we implement a machine learning technique to improve the resolution of the image. This technique learns
an end-to-end mapping between training data and corresponding target attributes.
1. Self-similarity based
Self-similarity based methods work on pairing low-resolution and high resolution patches to
improve the quality of low resolution image. Our method focuses on deep learning where we use
convolutional neural network and show that the deep learning based methods overcome the shortcomings
of state-of-the-art methods.
“The deeper the better” doesn’t work here. Our technique uses 3 convolutional-layer but doesn’t
show much of improvement for 4 to 5 layer. Thus, harnessing the power of GPU using less no of layers as
we have to consider training efficiency and storage.
We chose image super-resolution because of its vast applications in a number of fields like
forensics, image information enhancement, surveillance, medical diagnosis, earth-observation remote-
sensing, etc.
The low resolution images can be converted to high resolution without buying expensive
cameras.
It is one of the most active researches in the present scenario and also poses a lot of potential
observing its development over the past three decades.
1. Approximating high frequency information such as edges and texture that are lost.
2. Severely ill-posed inverse problems. There is no unique solution or they might not have
3. Solution in strict sense.
4. Algorithm should not exploit contextual information.
5. Most of the Super-Resolution algorithms work on single color channel. For colored images, we have to
work on three channels i.e. RGB which is to be converted to YCbCr.
2. LITERATURE SURVEY
According to the image priors, single-image super resolution algorithms can be categorized into
four types –prediction models, edge based methods, image statistical methods and patch based (or
example-based) methods. These methods have been thoroughly investigated and evaluated. Among them,
the example-based methods achieve the state-of-the-art performance.
The internal example-based methods exploit the self-similarity property and generate exemplar
patches from the input image and several improved variants are proposed to accelerate the
implementation. The external example-based methods learn a mapping between low/high-resolution
patches from external datasets. These studies vary on how to learn a compact dictionary or manifold
space to relate low/high-resolution patches, and on how representation schemes can be conducted in such
spaces.
Convolutional neural networks (CNN) date back decades and deep CNNs have recently shown an
explosive popularity partially due to its success in image classification . They have also been successfully
applied to other computer vision fields, such as object detection face recognition, and pedestrian
detection.Several factors are of central importance in this progress: (i) the efficient training
implementation on modern powerful GPUs,(ii) the proposal of the Rectified Linear Unit (ReLU) which
makes convergence much faster while still presents good quality, and (iii) the easy access to an
abundance of data (like ImageNet) for training larger models. Our method also benefits from these
progresses.
There have been a few studies of using deep learning techniques for image restoration. The multi-
layer perceptron (MLP), whose all layers are fully-connected (incontrast to convolutional), is applied for
natural image denoising and post-deblurring denoising. More closely related to our work, the
convolutional neural network is applied for natural image denoising and removing noisy patterns
(dirt/rain). These restoration problems are more or less denoising-driven. propose to embed auto-encoder
networks in their super-resolution pipeline under the notion internal example-based approach. The deep
model is not specifically designed to be an end-to-end solution, since each layer of the cascade requires
independent optimization of the self-similarity search process and the auto-encoder. On the contrary, the
proposed SRCNN optimizes an end-to-end mapping. Further, the SRCNN is faster at speed. It is not only
a quantitatively superior method, but also a practically useful one.
➢ ECONOMICAL FEASIBILITY
➢ TECHNICAL FEASIBILITY
➢ SOCIAL FEASIBILITY
ECONOMICAL FEASIBILITY
This study is carried out to check the economic impact that the system will have on the
organization. The amount of fund that the company can pour into the research and development of the
system is limited. The expenditures must be justified. Thus the developed system as well within the
budget and this was achieved because most of the technologies used are freely available. Only the
customized products had to be purchased.
TECHNICAL FEASIBILITY
This study is carried out to check the technical feasibility, that is, the technical requirements of
the system. Any system developed must not have a high demand on the available technical resources.
This will lead to high demands on the available technical resources. This will lead to high demands being
placed on the client. The developed system must have a modest requirement, as only minimal or null
changes are required for implementing this system.
SOCIAL FEASIBILITY
The aspect of study is to check the level of acceptance of the system by the user. This includes the
process training the wet to use the system efficiently. The user mast not feel threatened by the system,
instead must accept it as a necessary. The level of acceptance by the users solely depends on the methods
that are employed to educate the user about the system and to make him familiar with i
3. DESIGN
The requirements specifications for any system can be broadly stated as given below
4. IMPLEMENTATION
Our technique uses 3 convolutional-layer but doesn’t show much of improvement for 4 to 5 layer.
Thus, harnessing the power of GPU using less no of layers as we have to consider training efficiency and
storage. Traditionally, the image is up-scaled using bi-cubic interpolation. But it leads to loss of texture
due to excessive smoothness which leads to unnatural texture. Instead, we use discrete-wavelet transform
for upscaling and back-projection technique to enhance the edges. The up-sampled image is then
iteratively back projected using back-projection filter based on Laplacian of Gaussian (LOG) this gives a
more natural texture of the image in smooth areas. The method also removes spurious colors along the
edges.
Assumption:
Else:
Consider a single low-resolution image, we first upscale it to the desired size using bi-cubic interpolation,
which is the only pre-processing we perform. Let us denote the up-scaled image as Y.
We consider up-scaled image Y as input image on which we wish to learn a mapping F which
conceptually consists of 3 operations:
2. Non-Linear Mapping
3. Reconstruction
fig:An illustration of sparse-coding-based methods in the view of a convolutional neural network
We extract patches from the image which is a stratagy used for image restoration and apply below
operation,
We are having k1 dim feature for each and every patch.we will map this k1-dim vectors to k2-dim one
with filter size 1x1 by applying below operation,
b2 is k2 dimensional.
4.3 Reconstruct
By applying below operation we construct the final image