Analysis of Various Single Frame Super Resolution Techniques for better PSNR

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056
Volume: 03 Issue: 02 | Feb-2016 www.irjet.net p-ISSN: 2395-0072
© 2016, IRJET | Impact Factor value: 4.45 | ISO 9001:2008 Certified Journal | Page 1303
Analysis of Various Single Frame Super Resolution Techniques for
better PSNR
Sweta Patel1, Karshan Kandoriya2
1Student, Dept. of Computer Science and Engineering, Parul Institute of Engineering and Technology, Gujarat,
India
2Ass. Professor, Dept. Of Computer Science and Engineering, Parul Institute of Engineering and Technology,
Gujarat, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract – Single-frame Super-Resolution generate high
resolution image from single degraded image or low
resolution image. In single frame SR technique, the missing
high frequency information in the LR image is estimated from
large number of training set images and added to the LR
image. It require only one LR image for reconstruction so it
has more practical value for various applications. In this
paper, we are going to review different Super Resolution
methods which generate high resolution image from one or
more low resolution image.
Key Words: Spatial Resolution, Single frame, Super-
Resolution, Interpolation, Restoration
1. INTRODUCTION
In recent decades, Image Super-Resolution [1] is
broadly used research area and it solve the resolution
enhancement problem by quality of its optics, sensor
and display components. But, high resolution improve
by hardware is usually expensive and/or time
consuming. Therefore, we can increase the resolution
in two ways either by increase the pixel numbers or by
increase chip size. But it can degrade the acquisition
time and quality of image. So there is alternative
approach to enhance resolution of the image. Super
resolution is process of obtaining high qualityofimage
via single low resolution image or numerous. It can be
used in security surveillance, biomedical applications,
remote sensing, enlarging photograph for high quality
etc. Super resolution techniques can be classified into
twomajorparts:Frequencydomainandspatialdomain
approach. Frequency domain approach can perform
Fourier transform of an image. These methods have
low computational complexity, simple and more
suitableforremovingaliasingthanspatialdomain.This
method is popular but expensive. Spatial domain
approach allow more flexibility in incorporating a
priori constraints, noise models, and spatially varying
degradation models. Technical implementation of
super resolution canbedoneintwoways:single-frame
andmulti-frameSuper-Resolution.Single-framesuper-
resolution methods generate high resolution image
from single degraded noisy image.
There are manymethodsproposedfor singleimage
super resolution.Thesemethodscanbesummarizedas
methods based on reconstruction based approach and
learning based approaches. In this paper brief
description of these methods has been given.Insection
2 introduction of each method is given. Next, section 3
gives comparison of these methods. Final section
concludes the paper.
Fig-1: Example of Single image Super Resolution
1.2 Basic Architecture of Super Resolution
In Single-frame imaging [1], performing the observed
LR image from HR image is modeled by:
=D x + , k=1,…,n
Where is a warp matrix, including the global or
local translation, rotation etc., represents a blur
matrix, D is a sub sampling matrix, x is an original
image and represents a noise vector.
Figure 2 shows basic architecture for super resolution.
It consist of steps like estimation ofrelativemotion,i.e.,
registration, non-uniform interpolation, and de-
blurring. In Registration step to the estimation of the
relative shifts of each LR frame with respect to a

reference LR image with sub-pixel accuracy, Non-
uniform interpolation is to produce an improved
resolution image and image restoration is applied to
the up-sampled image to remove blurring and noise.
Before apply any technique on image, preprocessing is
carried out on image to remove noise or to convert
image into gray scale image. There are many
techniquesforsuperresolutionsuchasMAPestimation
approach, IBP estimation approach, Markov Network,
sparse representation, Manifold approach, etc.
Fig-2: Basic Steps of Super resolution
2. SUPER RESOLUTION METHODS
Super Resolution methods are mainly categorized in
reconstruction based approach and learning based
approach. These approaches are briefly described as
below:
2.1 Reconstruction based Approach
In Reconstructed based approach, incorporate the
prior knowledge to model a regularized cost function.
The image priors include the gradient prior, non-local
self-similarity and the sparsity priors. These prior
characterize different and complementary aspects of
natural image feature. Therefore combinational on of
multiple image prior for SR model may be beneficial to
improvement of performance.
Primal sketches Method
Primal sketches [3] were used as the a priori. The
primal sketches is the hallucination algorithm is
applied only to the primitives like edges, ridges,
corners, T-junctions, and terminations but not to the
non-primitive parts of the image. Having an LR input
image they first it to the target resolution, then for
every primitive point a 9×9 patch is considered. Then,
based on the primal sketch prior and using a Markov
chain inference, the corresponding HR patch for every
LR patch is found and replaced. This actually
hallucinates the high-frequency complement of the
primitives.
Gradient profile prior Method
A gradient profile-based [4] methods learn the
similarity between the shape statistics of the LR and
HR images in a training step. This learned information
will be used to apply a gradient based constraint to the
reconstruction process. The distribution of this
gradient profile prior is defined by a general
exponential family distribution.
IBP Estimation Approach
The IBP reconstruction [8]processistoestimateanHR
image as an initial solution first and then use
formula to calculate its simulative LR image
according to system model,
= +
If equals to the original HR image accurately, at the
same time when simulating the imaging process in
conform to the actual situation, then the simulated LR
image should be the same as the observation LR
image y, but when the twovectorsaredifferent,project
the error y- onto the reversely and
make correct further by using formula ,
= + (y- )
Where is HR image through the first time
revised, is a back projectionkernel.Thisprocessis
repeated iteratively to minimize the energy of the
error.

MAP Estimation Approach
Maximum A-Posteriori (MAP) approach [8], is popular
because it provides a flexible and convenient way to
include an a priori information and builds a strong
relationship between the LR images and the unknown
HR image. It is the basic idea of MAP estimation
approach to treat the HR image and observation LR
images as two different stochastic process, to regard
the SR reconstruction problem as a statistical
estimation problem, to make the HR image reaching
MAP on the premise of known LR image sequence. So
MAP estimator of the ideal HR image x is to maximizea
posteriori probability density function.
P (x| , , ..., )
Regularization Estimation Approach
Regularization method [8] is to use the solution of
prior information, convert ill-posed problems to well-
conditioned. Such as to minimize Lagrange equation
with constrained least square method by formula
min 2 + 2
Where C is a high-pass filter, which can eliminate the
influence of small singular values appearing in high
frequency, is referred to as the regularization
parameter,whichcanbalancetherelationshipbetween
smoothness and validity of the solution. A observation
model, which fuses together optical system prior
knowledge, use the iterative registration algorithm
based on gradient, and consider the optimization
process of gradient descent method and conjugate
gradient method to minimize cost function.
2.2 Learning based Approach
Learning-based or Hallucination algorithms were first
introduced in which a neural network was used to
improve the resolution of fingerprint images. These
algorithms contain a training step in which the
relationship between some HR examples (from a
specific class like face images, fingerprints, etc.) and
their LR counterparts is learned. This learned
knowledge is then incorporated into the a priori term
of the reconstruction. The training database of
learning-based SR algorithms needs to have a proper
generalization capability.
Markov Network based Approach
Markov network [8] to model the space relationship of
image from the perspective of probability for the first
time, divide image into small blocks, assume that each
image block is corresponding to a node on the Markov
network, and acquire transition probability matrix
for adjacent HR image block and transition probability
matrix Φ between HR image block and LR image block
through learning. The Markov network model can be
expressed by,
P(x ‫׀‬ y) = ( ) ( )
where y is input LR image, x is the HR image to
estimate, Z is the normalized parameter, and are
local adjacent image block in the HR image, is the LR
image block corresponding to
Manifold-based method
In Manifold-based approach [7], HR and LR images
form manifolds with similar local geometries in two
distinct feature spaces. This method is also used for
dimensionalityreduction.ItFinditsk-nearestneighbor
representation in low dimensional manifold for each
image blockin the testsamplesand usethesek-nearest
neighbors to calculate weighted coefficient and then
use the weights and neighbors to find the
corresponding objects in the HR manifold to
reconstruct the HR patch. In this situation, each HR
image block is related to its corresponding LR image
block which determine the reconstruction accuracy
and maintain some kind of connection among its
neighborhood block which decide the local preserving
feature and smoothness of the reconstruction image.
Sparse Representation Approach
In Sparse Representation [5] the signal is represented
by approximation of an image/signal with linear
combination of only small set of elementary signal
called atoms. Atoms are chosen eitherfrompredefined
set of function or learned from training set. The sparse
representation of high resolution image can be
recovered from low resolution image patches.Ituseda
small set of randomlychosen image patchesoftraining
and their SR method only applied to images with
similar statistical nature. To obtain dictionary use HR
image set for training HR dictionary. Correspondingly
the image of this set are then down sampled and

blurred and then used in training LR dictionary. Once
two dictionaries are obtained up-scaling of LR image
and calculate representation co-efficient using LR
dictionary by employing vector selection algorithm.
Then invoking assumption that representation co-
efficient of the desired HRpatch array are the sameHR
patch array is reconstructed.
Then the feature extraction is done using high pass
filtering operations.
Table -1: Comparison of various super resolution methods
3. COMPARATIVE STUDY OF DIFFERENT
METHODS
Table 1 shows the comparison among the different
methods used for the detection of defects in images.
Method Advantages Disadvantages
Primal sketches [3] It is use to enhance edges,
ridges and corner.
It require large databases of millions of high
resolution and low resolution patch pair and
are therefore computationally intensive.
Gradient profile prior [4] It is use gradient profiles in
natural images is robust
against changes.
Prior based on gradient profile need large set
of natural images.
IBP estimation approach
[8]
It is simple in principle and
easy to realize.
The response of the iteration need not always
converge to one of the achievable solution.
MAP estimation approach
[8]
It can join with expected
properties prior to minimize
cost function, and introduce
the priori regularity
constraints to ensure the
uniqueness of the solution.
Ability for preserving edges and details of the
reconstruction image is relatively weaker.
Markov network [8] It can obtain high frequency
information and high quality
image under the condition of
magnification 4 times.
Its efficiency is that request of training sample
choice is high and sensitive to noise.
Manifold-based approach [
7]
It require less training
sample with low sensitivity to
noise compared to Markov
network.
It is difficult to choose neighborhood block
size.
Sparse Representation
approach [5]
It overcome the problem of
choice of neighborhood size.
Generally, it is computationally intractable and
Solution is typically nonzero in every
component.

4. CONCLUSIONS
In this paper, survey of various methodologies for
super resolution is presented. These methods can be
classified into two categories: reconstruction and
learning based approach. A brief description of these
method including advantages and disadvantages is
given wherever known. So the combination of these
approaches can give better result instead of using
individual approach.
REFERENCES
[1] S. Park, M. Park, and M. G. Kang, “Super-resolution
image reconstruction: A technical overview,” IEEE Signal
Process. Mag., vol. 20, no. 3, pp. 21–36, May (2003).
[2] Z. Wang, A. C. brans.ovik, H. R. Sheikh andE.P.Simoncelli,
“Image quality assessment: From error visibility to
structural similarity, IEEE Image Process .,vol. 13, no. 4,
pp,600-612,2004.
[3] Sun, J., Zhang, N.N., Tao, H., Shum, H.Y.: Image
hallucination with primal sketch priors. ProceedingsofIEEE
Conference on Computer Vision and Pattern Recognition 2,
729–736 (2003).
[4] Sun, J., Sun, J., Xx, Z.B., Shum, H.Y.: Imagesuper-resolution
using gradient profile prior. In: Proceedings of IEEE
International Conference on Computer Vision and Pattern
Recognition, USA (2008).
[5] Yang, J., Wright, J., Huang, T., Ma, Y, ” Image super-
resolution via sparse representation “, IEEE ICIP, 2010.
[6] Kamal Nasrollahi, Thomas B. Moeslund: Super-
resolution: a comprehensive survey ,Springer Machine
Vision and Applications (2014) 25:1423–1468.
[7] Chang H, Yeung DY, Xiong Y. Super-resolution through
neighbor embedding. In: IEEE computer society conference
computer via pattern recognition; 2004. p. I-275–I-282.
[8] Lu Ziwei, Wu Chengdong1, Chen Dongyue,QiYuanchen1,
Wei Chunping: Overview on Image Super Resolution
Reconstruction, IEEE 2014.

Analysis of Various Single Frame Super Resolution Techniques for better PSNR

More Related Content

What's hot (16)

Similar to Analysis of Various Single Frame Super Resolution Techniques for better PSNR (20)

More from IRJET Journal (20)

Recently uploaded (20)

Analysis of Various Single Frame Super Resolution Techniques for better PSNR