Vol No. 1 Issue No.

1 International Journal of Interdisciplinary Engineering (IJIE) ISSN: 2456-5687

IMAGE SUPER-RESOLUTION VIA SPARSE

K.Surya Madhuri1 Prof.P.V.G.K. Jagannadha Raju2

ABSTRACT
To generate high resolution image from a low resolution input image single image super resolution is used. Single
image super resolution is used to enhance the quality of image. In this paper there is an image super resolution
algorithm is proposed which is based on GPS Gradient Profile Sharpness. Extract GPS from two gradient profile
description models.GPS contain two models one is triangle model and second is mixed Gaussian model. Finally the
HR(High Resolution)image is generated which has better visual Quality, lower resolution error. In this paper a new
algorithm for single-image super-resolution based on sparse representation over a set of coupled low and high
resolution dictionary pairs is proposed. The sharpness measure is defined via the magnitude of the gradient operator
and is shown to be approximately scale invariant for low and high resolution patch pairs. It is employed for clustering
low and high resolution patches in the training stage and for model selection in the reconstruction stage. A pair of
low and high resolution dictionaries is learned for each cluster. The sharpness measure of a low resolution patch is
used to select the appropriate cluster dictionary pair for reconstructing the high resolution counterpart. The sparse
representation coefficients of low and high resolution patches are assumed to be equal. By multiplying the high
resolution dictionary and the sparse coding coefficient of a low resolution patch, the corresponding high resolution
patch is reconstructed. Simulation results in terms of PSNR and SSIM and visual comparison, indicate the superior
performance of the proposed algorithm compared to the leading super-resolution algorithms in the literature over a
set of natural images in sharp edges and corners.

Index TermsSuper Resolution, Gradient profile, sparse representation; sharpness measure-based clustering;
multiple dictionary pairs

I.INTRODUCTION

The main aim is the generation of high resolution image from low resolution image .Image super resolution
process is widely used in some application. This application may be the image display, remote sensing and
satellite imaging[1].

Sometime take only one input image called as single image super resolution. In some of the Cases Multiple
input image may be considered the goal of conversion from low resolution image i.e. blurred image to better
quality image. Denoising and deblurring are the applications of image processing. When we construct high
resolution image from low resolution image there is one challenge that is the quality of edge should be
maintain. Super resolution is the technique of enhancement of image resolution. The pixel density within the
image is high that is the meaning of high resolution. To recover a high resolution image by using one or
more low resolution input images is the aim of super resolution methods. In this field previous work in this
field are classified into three categories. First is the Interpolation based approach ,second is the Learning

September 2016 Inside Journal (www.insidejournal.org) Page | 26

based approach and third is the reconstruction based approach. The main aim of image interpolation
approach is to reconstruct the edge of the original image. Interpolation is used to solve high dimension of set
of equation. Interpolation method is used to calculate the value of function at position between its sample it
reduce the signal bandwidth by using low pass filter to the digital signal .in image processing it is one of the
fundamental operation. Nearest neighbor, linear and spline interpolating techniques are the most commonly
used methods. Polynomial and language interpolation methods are less commonly used methods. The
fundamental operations in image processing are the process of interpolation. The quality of image is mostly
depending on the used interpolation technique. In interpolation process the information of edge direction is
very important. The interpolation direction is equal to the edge with the edge. By using sequence of images
to perform the interpolation super resolution interpolation can be improved. In learning based approach we
take low resolution image which is not clear by using this approach we compare this image to that image
which is already stored in dictionary. The content which are lost in low resolution image are retrieve from
that dictionary[1]. In Learning method with the help of training data set that predicate high frequency detail
lost in LR image. Learning method classified into 3 categories.

1) Nearest Neighbour Embedding Based SR method

In Reconstruction method it requires different patches from images then it resizes the SR output. Super
resolution is the process of converting LR image into HR image. Reconstruction based SR algorithm that
will require the patches of image from one or more images when synthesizing the SR output. Statistical
prediction model can be useful for single image super resolution. The model used to avoid any invariance
assumption. This model to further enhance performance that suggests data clustering. Image super
resolutions are technique aiming at resolution improvement of images acquitting by low resolution sensors,
minimizing visual effects. Recently several attempts have been made beyond the invariance assumption
aiming at improving the stability of recovery. The statistical model and the basic single image SR scheme:
The main motivation of this model to suggest the desire predict for each LR patch .The missing high
resolution detail This is mainly for two reasons first is the aims at characterizing signal of different quality,
so it use fewer atoms for the lower quality contend. Second is Use for avoiding high complexity sparse
coding computation. In this paper GPS is used to maintain gradient magnitude and spatial scattering of
gradient profile then two models that is triangle model and Gaussian mixture model is proposed. Profile
shape and profile gradient magnitude is maintained based on gradient profile transformation. Finally HR
image is generated from transformed gradient profile. That image added as an image prior in HR
reconstruction image.

II.RELATED WORK

Gradient profile is for describing the shape and sharpness of image when we convert low resolution image
into high resolution image the gradient profile is used for edge sharpness. Gradient profile prior is used for
improving the low resolution of image. For interpolation of the image the several information on the
sparsity of image has also useful. Reconstruct a high resolution image by using low resolution image is the
aim of image interpolation .For recovering sharp edges of low resolution image the image super resolution
will be use. When we convert the image from low resolution to high resolution it will be useful in many
application such as medical image diagnosis, computer vision, satellite imaging and also in entertainment.
Because of the imaging environment and the expensive imaging equipment when we capture the HR image

September 2016 Inside Journal (www.insidejournal.org) Page | 27

by using CCD and CMOS it is difficult to getting an image at the desired resolution level. So because of this
many super resolution methods are developed. Low resolution image is the down sampled and blurred
version of original HR image. To improve the resolution of given image SR technique are used. There are
two types of edges first is roof edge and ridge edge.

III.METHODS

When we are generating HR image it is very important to preserve the constraint of edge preservation in
interpolation algorithm.so for that purpose use gradient profile sharpness. To understand better description
of gradient profile two models are proposed first is Triangle model and second is mixed Gaussian model.
Triangle model is with short length and the mixed Gaussian model having heavy tails. This two models are
describing the gradient profile with both symmetric as well as asymmetric shapes.

3.1 Triangle Model: The Triangle model is useful for profile description. The triangle model having two
sides that are fitted separately using the extracted gradient profile points of each side [1]. mT (x) ={kdx + h
if the value >=0 ..(1) otherwise 0
In equation (1) mT(x):is the gradient magnitude of pixel x. dx:is the distance between the pixel x and profile
peak x0. K and h:are the slope intercept parameter of the linear function. Triangle model is most suitable for
the profile description in the kind of gradient profile. triangle model and gaussian mixed model are proposed
to realize gradient profile. triangle model and gaussian mixed model are proposed to describe different
profile shapes. triangle model is proposed for short gradient profile and mixed gaussian model for heavy
tailed gradient profile .

3.2 Gaussian Mixture Model: It is a combination of two Gaussian model. when edge are clear and
profile shapes is difficult.in that condition mixed Gaussian model is used. Gaussian mixed model has better
performance when gradient profile are symmetric. when there is large threshold set on gradient profile
length which is described by mixed Gaussian model. mixed gaussian model used to describe gradient profile
when gradient profile length is set.

3.3 Gradient Profile Sharpness: Gradient profile is used for edge smoothness. The work which is done
previously on single image super resolution can be divided into three categories first is interpolation based
second is learning based and third is reconstruction based. interpolation is fast and easy but blur high
frequency details. Edge sharpness is the important factor of image quality. Gradient field is used for reduce
the effects of noise while the edge will be enhanced. when we are generating HR image it is very important
to preserve the constraint of edge preservation in interpolation algorithm.so for that purpose use gradient
profile sharpness. Triangle model and Gaussian Mixed Model contain two feature first is there height h and
second is spatial scattering d based on two gradient profile description model metric of gradient profile
sharpness is defined. there for which is the ratio of height to spatial scattering.

n=h/d h=height represent the edge constraint which is the magnitude of gradient profile d=spatial scattering
represent edge of spatial spread. The edge contrast and edge spatial scattering this two point are GPS take
into consideration .the example of edge contrast is human perception of edge sharpness. edge sharpness are
well defined in GPS. Gradient profile is the advantage that describes spatial layouts of edge gradient
.triangle model and mixed gaussian model are used to specify gradient profile. By modeling edge gradient
profiles solve the image super resolution problem.

September 2016 Inside Journal (www.insidejournal.org) Page | 28

The original image is (a) and (b) is the edge sharpness image obtained from GPS. sharp edge and smooth
edge can be separate easily in image (b)

3.4 Gradient Profile Prior: Gradient Profile Prior can be used to propose image super resolution
approach in which shape and sharpness of image is described Gradient profile prior is used for gradient field
of natural image. on the basis of more prior information on the signal properties the sparse super resolution
algorithm can be enhanced[5].To achieve the better reconstruction quality the patch based approach are
basically work with maximally overlapping patches[6].Two widely used priors are generic smoothness prior
and edge smoothness prior. Able to provide a constraint on the gradient field of the HR image by using the
learned gradient profile prior and relationship. The recovery of high quality HR image by adding the
constraint of reconstruction. For image upsampling there are also contain edge statistics. For super
resolution, denoising and also for deblurring the sparseness prior can be successfully applied[12].How to
apply effective prior or constraint on HR image is the important issue for designing a better super resolution.

Advantages 1)small scale and large scale detail can be recovered in HR image because gradient profile prior
is not smoothness constraint 2)ringing fact can be avoided.

3.5 Gradient Profile Transformation(GPT): Using gradient profile prior they can approximate HR
gradient field by transforming LR gradient field this is done on the basis of gradient profile
transformation[12]. In this Section two models for GPT is proposed first is Gradient Profile Transformation
for triangle model and second is Gradient Profile Transformation for Gaussian mixture model. Three main
features are used to preserve the energy and shape of original Gradient Profile during Gradient Profile
Transformation 1)The sum of profiles gradient magnitude remains unchanged 2) Gradient Profile
Transformation keeps its peak position remains unchanged to avoid edge shifting 3)The shape of Gradient
Profile Transformation remains same or unchanged with its original Gradient Profile a)Gradient Profile
Transformation for triangle model: The GPT between two triangle model is simplified or easy because it's
shape is fixed. b)Gradient profile transformation for Gaussian mixture model: The GPT between Gaussian
mixture models is complicated because its shape is not fixed.

Algorithm1. The Proposed Single-Image Super-resolution

1. INPUT: A LR Test Image, Cluster Dictionary Pairs.

September 2016 Inside Journal (www.insidejournal.org) Page | 29

6. Divide the extracted features into overlapping patches and reshape them into vectors.
7. for Each LR patch do
8. Calculate the SM of the LR patch.
9. Determine the cluster this patch belongs to.
10. Sparsely code the features of the corresponding MR patch over the cluster LR dictionary.
11. Reconstruct the corresponding HR patch by right-multiplying the HR dictionary of the same cluster
with the sparse codes of the MR features.
12. end for
13. Merge overlapping patches to obtain a HR image estimate.

IV.CONCLUSION

A new super-resolution algorithm is proposed. This algorithm is based on sparse representation over
multiple coupled dictionary pairs. It employs scale-invariance of SM as a data clustering for training and
model selection in reconstruction. SM is used for clustering the training data and for each cluster a pair of
LR and HR dictionaries is learned. In the reconstruction stage, the SM value of a LR patch is used to select
the cluster it belongs to. Then, the dictionary pair of the selected cluster is used to reconstruct the unknown
HR patch. The proposed algorithm is shown to be superior to the algorithms that employ a single dictionary
pair. This is especially noticed for images with sharp edges and wide spread of SM values. Numerical and
visual tests conducted over a set of benchmark images reveal a competitive performance of the proposed
algorithm as compared to that of the state of-the-art algorithms.

REFERENCES

[1]. Qing Yan ,Student Member,Yi Xu,Member,Xiaokang Yang,Senior Member, and Truong Nguyen
Fellow,IEEESingle image super resolution Based on Gradient Profile Sharpness.
[2]. H. Hou and H. Andrews, Cubic splines for image interpolation and digital filtering, IEEE Transactions on
Acoustics, Speech and Signal Processing, vol. 26, no. 6, pp. 508517, 1978.
[3]. X. Zhang and X. Wu, Image interpolation by adaptive 2-d autoregressive modeling and soft-decision
estimation, IEEE Transactions on Image Processing, vol. 17, no. 6, pp. 887896, 2008.
[4]. F. Zhou, W. Yang, and Q. Liao, Interpolation-based image superresolution using multisurface fitting, IEEE
Transactions on Image Processing, vol. 21, no. 7, pp. 33123318, 2012.
[5]. S. Mallat and G. Yu, Super-resolution with sparse mixing estimators, IEEE Transactions on Image
Processing, vol. 19, no. 11, pp. 28892900, 2010.
[6]. W. Dong, L. Zhang, R. Lukac, and G. Shi, Sparse representation based image interpolation with nonlocal
autoregressive modeling, IEEE Transactions on Image Processing, vol. 22, no. 4, pp. 13821394, 2013.
[7]. A. Giachetti and N. Asuni, Real-time artifact-free image upscaling, IEEE Transactions on Image
Processing, vol. 20, no. 10, pp. 2760 2768, 2011.
[8]. K. Zhang, X. Gao, D. Tao, and X. Li, Multi-scale dictionary for single image super-resolution, in IEEE
Conference on Computer Vision and Pattern Recognition. IEEE, 2012, pp. 11141121.
[9]. D. Glasner, S. Bagon, and M. Irani, Super-resolution from a single image, in IEEE International Conference
on Computer Vision. IEEE,2009,pp.349-356.

September 2016 Inside Journal (www.insidejournal.org) Page | 30