A Novel Blind SR Method to Improve the Spatial Resolution of Real Life Video Sequences

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072
© 2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1577
A Novel Blind SR Method to Improve the Spatial Resolution of Real Life
Video Sequences
1W. RAJESH KUMAR, 2U. SATHISH KUMAR
(M.Tech.)1,Dept of CSE, Acharya Nagarjuna University, Guntur, AP, (522510), INDIA
Assistant Professor2,(,M.Tech,(Ph.d)),Dept of CSE, Acharya Nagarjuna University, Guntur, AP, (522510), INDIA
---------------------------------------------------------------------------***-----------------------------------------------------------------------
Abstract-In the real life video sequences super resolutions
become a complication because of the complex nature of the
motion fields. So to improve the efficiency of the video
sequences we proposed a method i.e. blind super resolution.
In this method the overall point spread function of the
imaging system, motion fields, and noise statistics are
unknown. The first step is to estimate the blur is non-
uniform interpolation SR method and then a multi-scale
process was performed on the estimated sequence. First the
estimation will starts on few emphasized edges and after
performing some steps gradually the iterations increased on
more edges. For the faster convergence we are performing
the estimation in the filter domain in the place of pixel
domain. To conserve the edges and fine details we are using
a high resolution frames by using a cost function that has
the fidelity and regularization terms of type Huber–Markov
random field. We performed masking operation to suppress
the artifacts which occurred due to inaccurate motions. By
using this operation the fidelity term is adaptively weighted
at each iteration. In the proposed method we got good
results for real-life videos containing detailed structures,
complex motions, fast-moving objects, deformable regions,
or severe brightness changes. We can see the results in both
subjective and objective evaluations.
Keywords: Video super resolution, blur deconvolution,
blind estimation, Huber Markov random field (HMRF).
1. INTRODUCTION:
Multi-frame super resolution, namely estimating the
higher frames from a low-res sequence, is one of the
fundamental problems in computer vision and has been
extensively studied for decades. The problem becomes
particularly interesting as high-definition devices such as
high definition television HDTV (1920 × 1080) dominate
the market. The resolution of various display has increased
dramatically recently, including the New iPad (2048 ×
1536), 2012 Macbook Pro (2880 × 1800), and ultra high
definition television UHDTV (3840 × 2048 or 4K, 7680 ×
4320 or 8k). As a result, there is a great need for
converting low resolution, low-quality videos into high-
resolution, noise free videos that can be pleasantly viewed
on these high resolution devices.
Although a lot of progress has been made in the past 30
years, super resolving real-world video sequences still
remains an open problem. Most of the previous work
assumes that the underlying motion has a simple
parametric form, and/or that the blur kernel and noise
levels are known. But in reality, the motion of objects and
cameras can be arbitrary, the video may be contaminated
with noise of unknown level, and motion blur and point
spread functions can lead to an unknown blur kernel.
Therefore, a practical super resolution system should
simultaneously estimate optical flow [3], noise level [7]
and blur kernel [4] in addition to reconstructing the higher
image. As each of these problems has been well studied in
computer vision, it is natural to combine all these
components in a single framework without making
simplified assumptions.
In this paper, we propose a Bayesian framework for
adaptive video super resolution that incorporates high-
resolution image reconstruction, optical flow, noise level
and blur kernel estimation. Using sparsity prior for the
high-res image, flow fields and blur kernel, we show that
super resolution computation is reduced to each
component problem when other factors are known, and
the MAP inference iterates between optical flow, noise
estimation, blur estimation and image reconstruction. As
shown in Figure 1 and later examples, our system
produces promising results on challenging real-world
sequences despite various noise levels and blur kernels,
accurately reconstructing both major structures and fine
texture details. In-depth experiments demonstrate that
our system outperforms the state-of-the-art super
resolution systems [1], [8], [9] on challenging real-world
sequences.
2. LITERATURE SURVEY
In this section we discuss a major class of super-resolution
methods which utilize a frequency domain formulation of
the super-resolution problem. The techniques discussed

utilize the shifting property of the Fourier transform to
model global translational scene motion, and take
advantage of the sampling theory to enable effect
restoration made possible by the availability of multiple
observation images. It is interesting to note that the
methods we discuss here include the earliest investigation
of the super-resolution problem, and although there are
significant disadvantages in the frequency domain
formulation, work has continued in this area until
relatively recently when spatial domain techniques, with
their increased flexibility, have become more prominent.
This does not however mean to say that frequency domain
techniques be ignored. Indeed, under the assumption of
global translational motion, frequency domain methods
are computationally highly attractive. We begin our review
of frequency domain methods with the seminal work of
Tsai and Huang [4].
In this paper [3], the author describes Printing from an
NTSC source and conversion of NTSC source material to
high-definition television (HDTV) format are some of the
recent applications that motivate super resolution (SR)
image and video reconstruction from lower resolution
(LR) and possibly blurred sources. Existing methods for SR
image reconstruction are limited by the assumptions that
the input LR images are sampled progressively, and that
the aperture time of the camera is zero, thus ignoring the
motion blur occurring during the aperture time. Because of
the observed adverse effects of these assumptions for
many common video sources, this paper proposes i) a
complete model of video acquisition with an arbitrary
input sampling lattice and a nonzero aperture time, and ii)
an algorithm based on this model using the theory of
projections onto convex sets to reconstruct SR still images
or video from an LR time sequence of images.
Kim et al [45] in their paper, propose a recursive algorithm
for restoration of super resolution images from noisy and
blurred observations. They use the aliasing relationship
between the under sampled frames and the reference
image, to develop a weighted recursive least-squares
theory based algorithm in the wave number domain. This
algorithm is efficient because interpolation and noise
removal are performed recursively and in addition, it is
highly suitable for implementation through the massively
parallel computational architectures currently available.
Accurate knowledge of the relative scene locations sensed
by each pixel in the observed images is necessary for super
resolution. This information is available in image regions
where local deformation can be described by some
parametric function. Such functions can describe, for
example, perspective transformation. Authors assumed
that local motion can be described by translations and
rotations only, but the approach is applicable also for other
image motion models.
3. PROPOSED METHOD
A. Observation Model
Fig 1: Estimation of fk frame from LR video
sequences
As shown in Fig. 1, a sliding window (temporal) of length
M + N + 1 (with M frames backward and N frames
forward) is overlaid around each LR frame gk of size
, and all LR frames inside the window are
combined through the SR process to generate the HR
reference frame fk of size . Here, Nx and Ny are
frame dimensions in two spatial directions and C is the
number of color channels. The linear forward imaging
model that illustrates the process of generating a LR frame
gi inside the window from the HR frame fk is given by:
( ) [ ( ( )) ( )]
( )
( )
where P is the total number of frames, (x↓, y↓) and (x, y)
indicate the pixel coordinates in LR and HR image planes
respectively, L is the down sampling factor or SR up scaling
ratio (so that , and is the two-
dimensional convolution operator. According to this
model, the HR frame fk is warped with the warping
function mk,i , blurred by the overall system PSF h, down
sampled by factor L, and finally corrupted by the additive
noise nk,i . It is more convenient to express this linear
process in the vector-matrix notion:
gi = DHMk,ifk + ni----------------(2)
Fig 2. Central motion (blue) versus sequential motion
(red).

In (2) fk is the kth HR frame in lexicographical notation
indicating a vector of size , matrices Mk,i and H
are the motion (warping) and convolution operators of
size , , D is the down sampling matrix of
size , and gi and ni are vectors of the ith
LR frame and noise respectively, both of size .
The matrix Mk,i registers (or motion compensates) the
reference frame fk to match the frame fi . As a result, Mkk is
an identity (unit) matrix since no motion compensation is
required between a HR frame and its coincident LR frame.
For a blur deconvolution (BD) problem (i.e. L = 1), D is the
identity matrix and so the input and output videos are of
the same size. Hence BD can be considered as a special
case of SR. The objective in SR and BD is to estimate the HR
frames fk and the blur H given the LR frames gi while the
motion Mk,i and the noise ni are unknown as well.
B. Color Space
The human visual system (HVS) is less sensitive to
chrominance (color) than to luminance (light intensity). In
the RGB (red, green, blue) color space, the three color
components have equal importance and so all are usually
stored or processed at the same resolution. But a more
efficient way to take the HVS perception into account is by
separating the luminance from the color information and
representing luma with higher resolution than chroma.
A popular way to achieve this separation is to use the
YCbCr color space where Y is the luma component
(computed as a weighted average of R, G, and B) and Cb
and Cr are the blue-difference and red-difference chroma
components. The YUV video format is commonly used by
video processing algorithms to describe video sequences
encoded using YCbCr.
In our implementation, video sequences can be processed
in either RGB or YUV formats. In the former case, SR is
used to increase the resolution of all R, G, and B channels.
However in the latter one, only the Y channel is processed
by SR for faster computation while the Cb and Cr channels
are simply up scaled to the resolution of the super-
resolved Y channel using a single-frame up sampling
method such as bilinear or bi cubic interpolation. The
obtained results related to these two cases are comparable
using a subjective quality assessment.
C. Motion Estimation
Accurate motion estimation (registration) with sub pixel
precision is crucial for video SR to achieve a good
performance. Two different approaches can be considered
for registration in video SR: central and sequential (Fig. 2).
In the former, motion is directly computed between each
reference frame and all LR frames inside its sliding
window (Fig. 1). By contrast, in the latter, each frame is
registered against its previous frame; then to use with SR,
sequential motion fields must be converted to central
fields for registration as follows: if Si = [ Sxi, Syi ] is the
sequential motion field for the ith frame (w.r.t. the (i −1)th
frame), then Mk,i = [ Mxk,i , Myk,i ] , the central motion field
for the ith frame when the central frame is the kth frame is
obtained as:
∑
∑
( )
Where I is the identity matrix.
With the sequential approach in SR, each frame needs to
be registered only against the previous frame, whereas
with the central approach each frame is registered against
all neighboring frames within its reconstruction window.
Therefore, the computational complexity and the storage
size of the motion fields in the central approach is higher
than that of using the sequential approach.
D. B LUR ESTIMATION
In a multi-channel BD problem, the blurs could be
estimated accurately along with the HR images [27].
However in a blind SR problem with a possibly different
blur for each frame, some ambiguity in the blur estimation
is inevitable due to the downsampling operation [19]. By
contrast, in a blind SR problem in which all blurs are
supposed to be identical or have gradual changes over
time, such an ambiguity can be avoided the assumption of
identical (or gradually changing) blurs makes it possible to
separate the registration and upsampling procedures from
the deblurring process which significantly decreases the
blur estimation complexity.
NUI method to reconstruct the upsampled frame
is explained. This upsampled yet-blurry frame is used to
estimate the PSF(s) and the deblurred frames through an
iterative alternative minimization (AM) process. The blur
and frame estimation procedures are discussed in Sections
III-B and III-C, respectively. The estimated frames are used
only for the deblurring process and so omitted thereafter.

Finally, the overall AM optimization process is described in
Section III-D.
E. Frame up sampling:
In [2] we discuss the situations in which the
warping and blurring operations in (2) are commutable.
Although for videos with arbitrary local motions this
commutability does not hold exactly for all pixels, however
we assume here that this is approximately satisfied. The
ultimate appropriateness of the approximation is validated
by the eventual performance of the algorithm that is
derived based on this model. With this assumption, (2) can
be rewritten as:
( )
Where = is the upsampled but still blurry frame.
Equation (4) suggests that we can first construct the
upsampled frames zk using an appropriate fusion method
and then apply a deblurring method to to estimate
and h.
If noise characteristics are also the same for all frames, an
appropriate way to estimate is using the NUI method
[3]–[5]. In NUI, the pixels of all LR frames are projected on
to the HR image grid according to their motion fields, and
then the intensities of the true locations on the grid are
computed via interpolation [2].
E. FRAME DEBLURRING:
After upsampling the frames, we use the following cost
function, J, to estimate the HR frames fk having an estimate
of the blur h (or H):
( ) ‖ ( )‖ ∑‖ ( )‖ ( )
where · 1 denotes the l1 norm (defined for a
sample vector x with elements xi as x = i | xi| ), λn is the
regularization coefficient, ρ(·) is the vector Huber function,
ρ(·) 1 is called the Huber norm, and ∇ j (i = 1,. . ., 4) are the
gradient operators in 0°, 45°, 90° and 135° spatial
directions [2]. The first term in (5) is called the fidelity
term which is the Hubernorm of error between the
observed and simulated LR frames. While in most works
the l2-norm is used for the fidelity term, we use the robust
Huber norm to better suppress the outliers resulting from
inaccurate registration. The next two terms in (5) are the
regularization terms which apply spatiotemporal
smoothness to the HR video frames while preserving the
edges.
Each element of the vector function ρ(·) is the
Huber function defined as:
( ) {
| |
| | | |
( )
The Huber function ρ(x) is a convex function that
has a quadratic form for values less than or equal to a
threshold T and a linear growth for values greater than T.
The Gibbs PDF of the Huber function is heavier in the tails
than a Gaussian. Consequently, edges in the frames are less
penalized with this prior than with a Gaussian (quadratic)
prior.
To minimize the cost function in (5), we use the
conjugate gradient (CG) iterative method [30] because of
its simplicity and efficiency. Compared to some other
iterative methods such as Gauss-Seidel (GS) or SOR that
need explicit derivation of matrix A when solving a linear
equation Ax = b, CG can decompose the matrix A to
concatenation of filtering and weighting operations.
However, CG can only be used with linear equation sets,
whereas the cost function in (5) is non quadratic and so its
derivative is nonlinear. To overcome this limitation, we
use lagged diffusivity fixed-point (FP) iterative method
[31] to lag the diffusive term by one iteration [15]. Using
this method for a sample vector x, at the nth iteration the
non-quadratic Huber-norm ρ(xn)1 is replaced by the
following quadratic form:
‖ ( )‖ ( ) ( ) ‖ ‖ ( )
Where Vn is the following diagonal matrix:
({
⁄
) (
In (8) the dots above the division and comparison
operators indicate element-wise operations. Applying the
FP method to (5) and setting the derivative of the cost
function with respect to fk to zero results in the following
linear equation set:
∑ ( )
Where

( ( ))
( ( )) (
We discuss how to update the regularization parameter λn
at each iteration in Section III-D.
F. Blur estimation:
Within an image or video frame, non-edge regions and
weak structures are not appropriate for blur estimation.
Hence, more accurate results would be obtained if the
estimation is not performed in such regions. For this
reason, in [11] and [33] the user should first manually
select a region with rich edge structure, whereas in [2],
[13], [14], and [34] the most salient edges are
automatically chosen. Moreover, sharpening salient edges
would also improve the accuracy of blur estimation. a
lowpass Gaussian filtering is utilized before shock filtering.
A similar concept for the blur estimation is exploited in
[37] in which the image is first sharpened by redistributing
the pixels along the edge profiles in such a way that ant
aliased step edges are produced. Having the sharpened
image and the blurry input image, the blur is then
estimated using a maximum a posteriori (MAP)
framework.
In our work, we employ the edge-preserving smoothing
method of [40] in which the number of surviving edges
after smoothing is globally controlled by the regularization
coefficient. This feature is helpful when one desires to limit
the number of salient edges at each iteration. This
smoothing method aims to keep an intended number of
non-zero gradients through l0 gradient minimization using
the following cost function:
( ) ‖ ‖ (‖ ‖ ‖ ‖ ) ( )
where fn k is the output of the edge-preserving
smoothing algorithm and the l0 norm is defined as x 0 = #
(i| xi = 0). Unlike shock filtering, this smoothing method
does not need pre-filtering of noise.
Though sufficient edge pixels are required for accurate
blur estimation, it is shown in [14] that structures with
scales smaller than the PSF support could harm blur
estimation. Inspired by that work, we define Rkn in (12) to
measure the usefulness of each pixel for blur estimation:
| | ( )
Where A and B are the convolution operators for
the spatial filters a and b, respectively, as defined below:
[ ] ( )
* + (
In (13) and (14), a is the all-ones filter of size 11 × 11 and b
is the sum-of-gradients filter. According to (12)-(14), to
compute , the sum of gradient components of
usually has a
small value at the location of narrow edges and smooth
regions.
𝑓0
𝑘
𝑧 𝑘
Algorithm 1 Blur Estimation Procedure
Require: 𝑔 𝑔 𝑝 𝜆 𝑚𝑖𝑛 Υ 𝑚𝑖𝑛 and initials 𝑜
𝜆 𝑜
𝛾 𝑜
𝛽 𝑜
𝑇 𝑜
𝑇2
0
Set n: =0 % Am loop iteration number
S: = # of scales
Use luma or one color channel of 𝑔 𝑔 𝑝1
for k: =1 to 𝑃 do % Loop on 𝑷 𝟏reference frames
if L>1 then % For SR reconstruction
𝑧 𝑘 𝑁𝑈𝐼(𝑔 𝑘 𝑀 𝑔 𝑘 𝑁)
else % For BD reconstruction
𝑧 𝑘 𝑔 𝑘
end if
%HR frame and blur estimation
for s: =1 to S do % Multi-scale approach
Rescale 𝑧 𝑘 𝑓𝑘
𝑛
𝑎𝑛𝑑 𝑛
% AM loop iteration
while “AM stopping criterion “is not satisfied do
n=n+1
% Updating procedure for f
compute 𝐕 𝐧
and 𝐖𝑗
𝑛
using (10)
update 𝜆 𝑛
while 𝐟 𝑛
does not satisfy “CG stopping criterion” do
𝑓𝑘
𝑛
: = CG iteration for system in (9); starting at
𝑓𝑘
𝑛
end while
Apply constraints on 𝑓𝑘
𝑛
% Updating procedure for 𝒉 𝒏
Update 𝛾 𝑛
𝛽 𝑛
𝑇 𝑛
𝑎𝑛𝑑 𝑇2
𝑛
Compute the smoothed frame 𝑓𝑘
𝑚
from (11)
Compute ∇𝑓𝑘
𝑛
from (15)
Edge tapping of ∇𝑓𝑘
𝑛
Compute 𝑘
𝑛
(𝑥 𝑦) from (17)
Apply constraints on 𝑛
end while
end for
end for
is
computed first, then at each pixel it is summed up with the
values of all neighboring pixels, and finally its absolute
value is obtained.
For pixels on narrow structures, the sum of gradient
values cancels out each other. Therefore,

Then is refined by only retaining strong and non-spike
edges:
{
| |
( )
Where and are threshold parameters which
decrease at each iteration.To avoid ringing artifact, we
apply the MATLAB function edgetaper() to ∇ . Then we
estimate each blur hk using the cost function J(h) below:
( ) ∑‖ ‖ ‖ ‖ ( )
Where P1 ≤ M + N and is the convolution
matrix of fk. Since J(h) in (16) is quadratic, it can be easily
minimized by pixel-wise division in the frequency domain
[41] as:
( )
(∑ ∑ {* ( ) ̇ ( )̇̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̇ ( ( ) ̇ ( )̇̇ ) +
̇
[| ( ) ̇ ( )̇ |
̇
| ( )| ̇
]}) ( )
Where ∇i(i = 1, 2) is ∇x or ∇y, F(·) and F−1(·) are
FFT and inverse-FFT operations, and (·) is the complex
conjugate operator. We then apply the following
constraints to the estimated PSF: its negative values are set
to zero, then the PSF is normalized to the range [0, 1], and
centered in its support window.
D. Overall Optimization for Blur Estimation:
The overall optimization procedure for estimating
the PSF is shown in Algorithm 1. The HR frames and the
PSF are sequentially updated within the AM iterations. We
use a multi-scale approach to avoid trapping in local
minima. The regularization coefficients in (9) and in
(17) decrease at each AM (alternating minimization)
iteration up to some minimum values and ,
respectively (see [2] for a discussion). The variation of
these coefficients is given by:
( ) ( ) ( )
Where r is a scalar less than 1. Also the values of
βn in (11) and Tn 1 and T2n in (15) fall at each AM
iteration which increases the number of contributing
pixels to blur estimation as the optimization proceeds.
IV. FINAL HR FRAME ESTIMATION
After the PSF estimation is completed, the final HR
frames are reconstructed through minimizing the
following cost function:
( ) ∑ ( ∑ ‖ ( ( ))‖
∑‖ ( )‖ ( ))
Where Ok,i is a diagonal weighting matrix that
assigns less weights to the outliers. Minimizing this cost
function with respect to fk yields:
Algorithm 2 Final Frame Estimation Procedure
Require: 𝑔 𝑔 𝑝 and 𝜆
1: Set n: = 0 % FP loop iteration number
2: for k: =1 to P do % Loop on P reference
frames
3: Estimate sequential motion fields 𝑆 𝑆 𝑝
4:Compute central motion fields 𝑀 𝑀 𝑝 using
(??)
5:Estimate the blur h using Algorithm 1
7:% Estimate HR frames using FP loops
8:while ‘’FP stopping criterion” is not satisfied do
9: n=n+1
10: Compute 𝐎 𝑘𝑗
𝑛
using (22)
11: Compute 𝐕 𝑛
and 𝐖𝑗
𝑛
using (21)
12: While 𝑓 𝑛
does not satisfy “”CG stopping
criterion” do
13:𝑓𝑘
𝑛
: = CG iteration for system in (20); starting at
𝑓𝑘
𝑛
14: end while
15 Apply constraints on 𝑓𝑘
𝑛
17: end while
18: end while

( ∑ ∑ )
( )
where
( ( ))
( ( )) ( )
and the m-th diagonal element of Onk,i is computed
according to:
[ ] {
‖ ( ( ))‖
} ( )
Where in (22) Rm is a patch operator which extracts a
patch of size q × q centered at the m-th pixel of fk,i .
The final frame estimation procedure is demonstrated in
Algorithm 2.
5. Experimental results
In this section, the performance of our method is evaluated
and compared with the state-of-the-art video SR methods
3D-ISKR 4 [38] and Fast Upsampler[42] which are
available for public evaluation, and also with the
commercial software Video Enhancer[43]. Among these
three, we only display the results from 3D-ISKR[38]. This
non-blind SR method does not include a deblurring step, so
we post process its outputs with the deblurring method of
[39]. Different parameters for deblurring were tried out in
each experiment to get the best possible outcomes from
3D-ISKR. Furthermore, since 3D-ISKR implementation
does not estimate pixels near frame boundaries, we
remove the boundaries from the reconstructed frame
before an objective evaluation.
Fig 1: Proposed method video sequence
Fig 2: degraded video sequence
Fig 3: Motion estimation in video sequence

Fig 6: Comparison results
Fig 4: inverse video sequence
Fig 5: Bi cubic interpolation pervious method Fig 7: PSNR values for proposed and pervious method
CONCLUSION:
In this paper by using the blind deconvolution technique
the low resolution video is converted into a super
resolution form. The main problem in the real-life video
sequences is in the nature of motion the frame will become
blur. By using the non-uniform interpolation (NUI) SR
method the input frames are upsampled and the blur is
estimated. In the estimation process an assumption is
made as the blur is either identical or have a slow
variations over time. After performing upsampling of
frames the blur is determined for some enhanced edges.
To obtain the final reconstructed frames the previous
frames are removed and non-blind iterative SR process is
performed on the estimated blur (s). To get the perfect
resolution masking operation is performed during the
each iteration of the final frame reconstruction. Proposed
method is compared with the conventional methods to see
the accuracy of the output performance.
REFERENCES:
[1] http://www.infognition.com/videoenhancer/, Sep.
2010. Version 1.9.5.
[2] S. Baker and T. Kanade. Limits on super-resolution and
how to break them. IEEE Transaction on Pattern Analysis
Machine Intelligence, 24(9):1167–1183, 2002.
[3] B.K.P. Horn and B.G. Schunck. Determining optical flow.
Artificial Intelligence, 16:185–203, Aug. 1981.
[4] D. Kundur and D. Hatzinakos. Blind image
deconvolution. IEEE Signal Processing Magazine, 13(3):43
–64, 1996.
[5] Z. Lin and H-Y +Shum. Fundamental limits of
reconstruction based superresolution algorithms under

local translation. IEEE Transaction on Pattern Analysis
Machine Intelligence, 26:83 –97, 2004.
[6] C. Liu and D. Sun. A bayesian approach to adaptive
video super resolution. In IEEE International Conference
on Computer Vision and Pattern Recognition, 2011.
[7] C. Liu, R. Szeliski, S. B. Kang, C. L. Zitnick, and W. T.
Freeman. Automatic estimation and removal of noise from
a single image. IEEE Transaction on Pattern Analysis
Machine Intelligence, 30(2):299–314, 2008.
BIOGRAPHY:
W. RAJESH KUMAR received the
Bachelor of Technology degree in
Computer Science engineering from
KL University in 2015. Currently
pursuing M.tech in Computer
Science Engineering from Acharya
Nagarjuna University. His areas of
interest include computer science
engineering specialized in image processing and Data
mining.
Mr. U.SATHISH KUMAR is currently
working as assistant professor in
Acharya Nagarjuna University from
last 3 years. Currently he is pursuing
his Ph.D degree. His research
interest include Compute Science
and Engineering in Data mining, Soft
computing and image processing

A Novel Blind SR Method to Improve the Spatial Resolution of Real Life Video Sequences

More Related Content

What's hot (18)

Similar to A Novel Blind SR Method to Improve the Spatial Resolution of Real Life Video Sequences (20)

More from IRJET Journal (20)

Recently uploaded (20)

A Novel Blind SR Method to Improve the Spatial Resolution of Real Life Video Sequences