IRJET- Exploring Image Super Resolution Techniques

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 06 | June 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 2889
Exploring Image Super Resolution Techniques
Varsha R1, Vibush Shanmugam2, Neha Ghaty3
1,2,3Student, Dept of CSE, PES University, Bangalore, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - Super resolution is a technique that
reconstructs a high resolution image from the observed low
resolution images. Most Super Resolution techniques aim to
improve the spatial resolution of an image. But as two-
dimensional signal records, digital images with a higher
resolution are always desirable in most applications.
Imaging techniques have been rapidly developed in the last
decades, and the resolution has reached a new level. The
question we are trying to address is whether image
resolution enhancement techniques are still required and if
they are, what are the techniques that are state of the art,
what type of super resolution enhancements are known to
produce the best results, what are the drawbacks with these
architectures and to come up with a work around to handle
these drawbacks and improve the resolution of images
Key Words: Computer vision, high resolution, super
resolution, spatial resolution, image resolution
1. INTRODUCTION
The main objective of super-resolution is to estimate the
high-resolution visual output of a corresponding low
resolution visual input, which can either be a low-
resolution image (single-image) or a set of images (multi-
image), for example, corresponding to frames in a video
sequence. The goals range from providing better content
visualization for traditional image processing application
to achieving better visual recognition, including computer
vision tasks. Image super-resolution is important in many
applications of multimedia, such as playing a video on a
higher-resolution screen. Due to some technical
limitations in imaging devices and systems, like, the
presence of optical distortions and lens blur, insufficient
sensor sampling density and aliasing, motion blur due to
low shutter speed, the presence of noise due to sensor
limitations and lossy coding, super-resolution technique is
actually needed. The high-resolution visual output can be
obtained either by providing devices with excellent spatial
resolution, at the cost of a very high market price of the
imaging device or with the use of software-related tools.
The former is achieved by some hardware-related tools
which includes – reducing the pixel size (which
unfortunately leads to an increasing appearance of shot
noise as the amount of light captured by the device
decreases), increasing the chip size to accommodate a
larger number of pixel sensors (which unfortunately
results in an increased capacitance), reducing the shutter
speed (which leads to an increasing noise level), adoption
of high-precision optics and sensors (which invariably
results in an increase in the price of the device). The
advantage of post-processing the captured visual data is
that it allows us to balance computational and hardware
costs. Thus, on one hand we may have a lower market
price and, on the other we can work with contemporary
imaging devices and systems.
Super-resolution allows a high-resolution image to be
generated from a lower resolution image, with the trained
model inferring photo-realistic details while up-sampling.
In this work, we will explore super resolution GANs and
their applications in detail.”
Super-resolution GAN applies a deep network in
combination with an adversary network to produce higher
resolution images. SRGAN is more appealing to a human
with more details. During the training, a high-resolution
image is downsampled to a low-resolution image. A GAN
generator upsamples the low resolution images to super-
resolution images. We use a discriminator to distinguish
between the original high resolution images and the super
resolution image generated by SRGAN. The GAN loss is
then backpropagated to train the discriminator and the
generator. The SRGAN model [9] adds an adversarial loss
component which constrains images to look like natural
images, producing convincing solutions.
2. LITERATURE SURVEY
2.1 Image super-resolution
The task of estimating a high-resolution image from its
low-resolution counterpart is referred to as super-
resolution (SR). The optimization target of super-
resolution algorithms is usually the minimization of the
Mean Square Error (MSE) between the generated image
and the ground truth image. Minimizing MSE also
maximizes Peak Signal to Noise Ratio (PSNR) which is a
common measure that is used to evaluate super resolution
algorithms.”
2.2 History of super-resolution techniques
In 1964, Harris established the foundation for super
resolution as a technique by solving the diffraction
problem [19]. The milestones of spectroscopy have been
achieved almost entirely by using readily available
detection technology while minimizing background levels.
The lesson from the progress in both fields is basically that
anything can be detected if the background is low enough.
In 1984, Tsai and Huang first addressed the idea of super
resolution to improve the spatial resolution of a dataset
containing the Landsat images. After analysing the results

from these experiments the super resolution techniques
were categorized into Interpolation based methods,
Reconstruction based methods and Experiment based
methods.
In the period of 1984 to 2000, most methods concentrated
on frequency domain based super resolution technique.
This technique comes under reconstruction methods to
obtain high resolution images that obtains high
computational efficiency. But it was observed that these
models were sensitive to errors and could not handle
complicated inputs.
2000 to 2010 was a decade of spatial domain methods.
Most of these methods produced state of the art results in
that era. But however now, these methods are obsolete
because of the advent of experiment based techniques.
Interpolation was one of the easiest and common
methods. Iterative back projection, regularization, etc
were a few other methods that were designed for super
resolution.
From 2010 until the present days, machine learning and
deep learning methods have changed the way we used to
solve problems. Computer vision has revolutionized the
image processing domain. Example based methods were
widely popularized and regression based methods, SR-
CNN, SR-GANs were the commonly used methods.
2.3 Comparison
1. Interpolation based methods: Interpolation is the
technique of using points with known values or sample
points to estimate values at other unknown points.
Advantages: Needs lesser computational complexity and
hence these methods are better suited for real-time
applications.
Disadvantages: It’s not possible to obtain high-frequency
information and it is not possible to find missing spectral
contents. The results obtained are over-smooth and have
jagged artifacts at edges.
2. Reconstruction based methods: Data collected in two-
dimensional projections give planar images of object at
each projection angle. To obtain information along the
depth of the object, tomographic images are reconstructed
using these projections.
Advantages: The high resolution images are very close to
the original low resolution images with respect to
features. There is also an added smoothness because of
down-sampling.
Disadvantages: Not suited for arbitrary images as ringing
artifacts may appear.
3. Example based methods: The algorithms of example-
based super-resolution problems are based on machine
learning models exploiting available examples.
Advantages: These are the most successful in producing
state of the art, best quality images because it is based on
example learning neural networks.
Disadvantages: Due to insufficient training examples
available for the model to learn from, high frequency
artifacts may appear in the output high resolution images.
Learning time also significantly increases which forces the
need for hardware resources like high memory or a GPU.
3. RESEARCH RESULTS
3.1 Interpolation based SR techniques
1. Nearest Neighbour Interpolation: This method
manipulates pixel values of the nearest pixels which have
the same value as the neighbour pixel. This method is one
of the simplest and easiest but does not produce hig sub-
pixel accuracy.
2. Bilinear Interpolation: This method passes a straight
line between two consecutive pixel locations. This method
is known to be better than Nearest Neighbour
Interpolation but still does create artifacts and poor
preservation image details.
3. Quadratic Interpolation: This method uses three points
for interpolation and results in one point at the centre and
another two points on each side. This has shown one of
the best performances.
4. Bicubic Interpolation: This method extends into four
number of pixel neighbours where the function is defined
with two pixels on each side. This performs better than
quadratic interpolation too.
3.2 Reconstruction based SR techniques
1. Non-uniform interpolation: This method allows for the
reconstruction of samples from other samples taken at
non-uniformly distributed locations. It is a basic and
intuitive method of super resolution and is known to have
relatively low computational complexity. But it assumes
that the blur and noise characteristics are identical across
all low resolution images.
2. Frequency domain: This method is basically
reconstructing a high-resolution image from multiple low
resolution images based on the aliasing images present in
the low resolution images. This method is simple to
implement and produces high quality output but it is only
efficiently applied provided the noise has zero mean the
blurring is either absent or identical across the low
resolution images.
3. Regularization: In this method, we assume that the
registration parameters are estimated and deterministic
regularization is done by taking proper prior information
about the solution. There is no need of larger training
datasets as the image details preservations is high. But the
performance degrades with higher magnification factor. It
also takes more time for computation.

4. Projection onto Convex Sets: In this method, an
estimate of the high resolution version of the reference
image is determined iteratively starting from some
arbitrary initialization. It solves the effect of under
sampling but suffers from image blurring.
5. Iterative back projection: In this image, a high
resolution image is estimated by back projecting the
difference between the simulated low resolution image
and captured low resolution image on the interpolated
image. This method removes noise and blurry effects from
the image but there is no unique solution as it is difficult to
choose the ideal back projecting operator.
3.3 Example based SR techniques
1. Neighbour Embedding: In this method, each input data
vector can be described as a linear combination of its
nearest neighbours on the natural image manifold of low
resolution patches. This is an unsupervised learning
method with both external and internal learning and
where external performs better.
2. Sparse Coding: This follows the previous method with
the additional constraint of a compact and optimized
dictionary that is obtained through the training process.
This provides robust nearest neighbour decomposition.
This is also unsupervised on both the low and high
resolution pairs. This allows only external learning. Also
ensures no overlapping and requires low computation.
3. Anchored Regression: In this method, an external
database composed of a low resolution dictionary and a
set of linear regression matrices map the low
resolution examples to their high resolution counterparts.
The running cost and computational cost is greatly
reduced by removing the sparsity constraint from the
inference stage in sparse coding technique. This allows
external learning methods.
4. Regression Trees and Forests: In this method, the input
image is divided into patches. Each patch traverses the
tree from root node to the most suitable leaf node, and the
corresponding regression model is used to generate the
high-resolution patch. This method is computationally
faster than other example based techniques. But there is a
limitation of high memory requirements for storing the
regression parameters because they are expanded in the
whole set of training data points.
5. Deep Learning: In this method, we use the power of
back-propagation algorithms in order to learn the
hierarchical representations that allow for minimizing the
error at the end of the network. Deep learning is one of the
current alternatives with supervised learning approach
based on deep convolutional neural networks. These
algorithms have the power to determine the hierarchical
descriptions of the visual data and this is learned directly
from the data. But the fine tuning of all the parameters in
the network takes a considerably large amount of time
than the classical machine learning approaches.
3.4 Challenges faced today
1. Image registration: Bayesian approach can be used but
computation cost can be very high.
2. Computational efficiency: Interpolation restoration
algorithms work but computation goes up with non-
translation models.
3. Robustness aspects: Median estimation to combine the
upsampled images to cope with outliers from noise works
but showed improvements for outliers assumed on the
validation data and not much on real data.
4. Performance limits: motion estimation, decimation
factor, number of frames and prior information work but
they only suggest ways and are far from enough.
3.5 GAN based SR techniques
Generative Adversarial Networks also commonly known
as GANs are deep neural-network architecture comprising
of the generator and discriminator, that are pitted against
each other . GANs have large potential since they can learn
to generate by itself any type of data.
[1]
The SRGAN model proposed by Ledig et al. [9] adds an
adversarial loss component to the GAN loss function which
constrains images to appear close to natural images. The
SRGAN generator is conditioned on low-resolution input
and infers photo-realistic natural images with 4x
upsampling. Along with adversarial loss there is a
perceptual loss (which is a weighted sum of content loss
and adversarial loss) from a pre trained classifier and
regularization loss that encourages spatially coherent
images. SRGAN set a new state-of-the-art for image super-
resolution with high upscaling factors (4x) as measured by
both PSNR and Structural Similarity (SSIM). They confirm
with an extensive mean opinion score (MOS) test on
images from three public benchmark datasets (Set5, Set14
and BSD100) that SRGAN is the new state of the art, by a
large margin, for the estimation of photo-realistic SR
images with high upscaling factors (4×) .It also shows that
perceptual loss is more invariant to changes in pixel space
and hence performs better than MSE based content loss.
Despite this SRGAN is not optimized for video super-
resolution in real time and perceptually convincing
reconstruction of text or structures scenes is still beyond
the scope of the model.”
Liu et al. [10] attempted to further improve on the SRGAN
model proposed by Ledig et al. by making changes to the
model network. They trained models for both 2x and 4x
upsampling of images. They tried three different loss
function/optimizer combinations, namely, softmax cross
entropy loss with an Adam optimizer, which did not train
beyond a few hours until either the generator or

discriminator gained an advantage over the other.
Wasserstein GAN with gradient penalties did not produce
useful results either. Finally, using the loss function from
Least-Squares GAN: least square loss and Adam optimizer
resulted in the most stable training and best-looking
images of the three approaches. They received results that
were better than bicubic interpolation methods,
sometimes at the expense of added background colour
noise and artifacts and some preliminary results on video
super-resolution. The model however had a hard time
generating details across all classes when the input
dataset had many class labels. They trained their 4x
upsampling GAN only on anime images and hence this
model may not work well on real images. Their
preliminary work on video super resolution led to a
generative network that could learn color and spatial
structure well, however there was still a little bit of blur.”
Xiangyu Xu et al. [11] created an algorithm to directly
restore a clear high-resolution image from a blurry low-
resolution input. They focus on text and face images and
learn a category specific prior to solve this problem. They
designed two models called MCGAN (Multi-class GAN) and
SCGAN (Single-class GAN) in order to generate high
resolution images. SCGAN has a single generator and
discriminator whereas MCGAN has a single generator and
K discriminators which are trained to classify real and
generated images for each of the K classes. After training
the learned generator can be used to generate images
from any of the K classes. They added two additional
components to the basic GAN loss function, namely pixel-
wise loss and a feature matching loss term. The pixel-wise
loss penalizes the difference between generated images
and the ground truth image. The feature matching loss
function forces restored and real images to have similar
feature responses at the intermediate layers of the
discriminator network. This creates more realistic
features and structural information in generated images.
Results showed that SCGAN performed better than
MCGAN and state-of-the-art super resolution methods on
both text and face images. On the down side some of the
reconstructed faces contained checkerboard artifacts.
Karras et al. [12] design a model that progressively trains
the generator and the discriminator. Generation of high
resolution images is difficult due to the gradient problem
and memory constraints. Progressively growing the
generator and the discriminator starting from low
resolution images and progressively adding higher
resolution images not only improves the stability in the
high resolution images generated, but also significantly
reduces the training time. GANs have a tendency to
capture only a part of the variation in the training data
which hampers high resolution image generation. They
proposed a new way to increase variation in the generated
data which uses minibatch (Salimans et al. [13]) standard
deviation to produce a new feature map. This layer is then
inserted towards the end of the discriminator to produce
optimal results. To disallow the scenario where the
magnitudes in the generator and discriminator spiral out
of control as a result of competition, they normalize the
feature vector in each pixel to unit length in the generator
after each convolutional layer. In addition, they also
introduce a new metric for evaluating quality and
variation of generated images, which uses sliced
Wasserstein distance (SWD). This is because current
techniques such as MS-SSIM (Odena et al., 2017) are good
at finding large-scale mode collapses easily but fail to react
to smaller effects such as loss of variation in colors or
textures. The experimental results achieved state of the art
inception scores of 8.80 for the CIFAR10 dataset and also
created a high-quality version of the existing CELEBA
dataset consisting of 30000 of the images at 1024 × 1024
resolution. However, they maintain that there is still a long
way to true photorealism and room for improvement in
the micro-structure of the images.”
Andrew Beers et al. [14] took this concept of PG-GANs a
little further and used them specifically for medical image
synthesis and the resolution of medical image data. The
same training scheme of progressively growing of GANs
has been used to create photorealistic and phenotypically
diverse fine-grained images at high resolution. The GAN is
trained in phases with each phase adding an upsampling
layer and a pair of convolutional layers to both the
discriminator and generator using an Adam optimizer and
the loss calculated as Wasserstein Loss. In this experiment,
High quality and variation images were produced
including some very unrealistic images. They did suffer
from some overly distinct edges mostly caused by the
pressure on the generator to create segmentation maps
but importance was given to certain pathological features.
PGGAN had the ability to produce a great variety of images
and the ability to generate vessel trees from outside its
original training set. It was also observed that the latent
space of GANs often encodes semantic information about
the images produced, and that latent vectors similar to
each other in latent space produce qualitatively similar
output images. This method can produce images of
unprecedented size and its latent space can be used to
learn imaging features in an unsupervised manner for high
resolution imaging.”
Huikai et al. [15] developed a high resolution conditional
image framework called GP-GAN that uses both GANs and
gradient based image blending methods. They build a
network called Blending GAN for generating low-
resolution realistic images by proposing the Gaussian-
Poisson equation to combine gradient information and
colour information. Blending GAN leverages Wasserstein
GAN for supervised learning tasks using the encoder-
decoder architecture. Gaussian-Poisson equation
fashioned by the well known Laplacian Pyramid is
proposed to make use of the natural images produced by
Blending GAN to generate high resolution realistic image
by approximating the color. The conditional GAN is good

at generating natural images from a particular distribution
but weak in capturing high frequency image details like
textures and edges. Gradient based methods on the other
hand perform well at generating high-resolution images
with local consistency but generated images tend to be
unnatural and have many artifacts. Combined together,
these methods could result in a conditional image
generation system and also for image-to-image translation
tasks. The drawback of the algorithm is that it fails to
generate realistic images when the composited images are
far away from the distribution of the training dataset.
In order to boost network convergence and achieve good
looking high resolution results, Curt´o et al. [16] proposed
a model called HDCGAN, with the goal being to generate
indistinguishable samples to push GANs to scale well and
maintain context information of high resolution images.
They create an image generation tool that samples from a
very precise distribution whose instances resemble or
highly correlate with real sample images of the underlying
true distribution. These generated image points fit well
into the originals and add additional information such as
redundancy, poses or generate highly-probable scenarios.
Self-normalizing Neural Networks (SNNs) keep the
activations normalized when propagating through the
layers and Scaled Exponential Linear Units (SELU) is used
as the activation function for feed forward neural
networks to construct a mapping leading to SNNs. SELU +
Batchnorm is used when numerical points move away
from the usual point. BatchNorm ensures it is close to a
desired value thus maintaining convergence. This
technique stabilizes training, allows fewer GPU resources
with steady diminishing errors in the generator and
discriminator thus accelerating convergence speed. High-
Resolution Deep Convolutional Generative Adversarial
Networks (HDCGAN) by stacking SELU + BatchNorm (BS)
layers generates high-resolution images in circumstances
where all other former methods fail. It exhibits a steady
and smooth training mechanism.
Bousmalis et al. [6] use unsupervised learning to learn a
transformation in the pixel space from one domain to
another. The model makes source-domain images appear
as if they are drawn from the target domain. They train a
model to change images from the source to appear as if
they are from the target domain, while maintaining the
original content and this is known as pixel level domain
adaptation method or PixelDA. In their model they make
use of content similarity loss that penalizes the difference
between source and generated image for foreground
pixels. They also use pairwise mean squared error (PMSE)
which penalizes the difference between pairs of pixels
rather than absolute difference between input and output.
The different domain adaptation scenarios that were
considered are MNIST to USPS, MNIST to MNIST-M and
Synthetic Cropped LineMod to Cropped LineMod. They
perform a quantitative and qualitative evaluation on the
different domain adaptation scenarios. qualitative
evaluation involves the examination of the ability of the
method to learn the underlying pixel adaptation process
from the source to the target domain by visually
inspecting the generated images. Qualitative evaluation
involves comparison of PixelDA with Source only(train on
source training data and evaluate on target test data) and
Target only(train on target training data and evaluate on
target test data) baselines. The PixelDA models
outperforms previous work on a set of unsupervised
domain adaptation scenarios, and in the case of the
challenging “Synthetic Cropped Linemod to Cropped
Linemod” scenario, the model more than halves the error
for pose estimation
Phillip et al. [17] investigate conditional GANs as a general
purpose solution for image to image translation problems
because they learn a loss that adapts to the data. In other
words, they learn a structured loss which penalizes the
entire output instead of treating each output pixel
independently. Conditional GANs learn a mapping from
the combination of observed image and random noise
vector to output image.The noise is provided in the form of
dropout to the generator during both training and test
time. In image translation, a lot of low level information is
common between the input and output images. In order to
move this directly across the network, skip connections
are added to the generator which follow the shape of a U-
Net [18]. The discriminator is a PatchGAN which penalizes
the image structure at the patch level. The discriminator
predicts real or fake images by averaging all the responses.
The model is evaluated on toe metrics, AMT perceptual
studies which involves testing plausibility to a human
observer and FCN-score which uses an off the shelf
classifier to see how well the synthesized images can be
classified. The results showed that using GANs along with
L1 loss function fooled participants 18.9% which is
significantly higher than previous methods. However
designing conditional GANs that produce highly stochastic
output which capture the full entropy of the conditional
distributions they model is an important question left
open by the present work.”
3.6 Analysis
Based on our analysis of various super resolution
techniques, we find that a lot of the challenges in super
resolution have been addressed and rectified. However
there are still some challenges or limitations that need to
be addressed.
1. Noise and motion blur
2. Image quality preservation
3. Oversmoothing in images
Interpolation, although simple to implement leaves much
to be desired in terms of visual quality, as the details (eg:-
sharp edges) are often not preserved. Sparse coding
makes use of low resolution and high resolution images,

however the pipeline involves multiple steps all of which
can not be optimized.
4. EXPERIMENTS
In an attempt to understand the performance of the
various super resolution techniques, we chose different
architectures to measure the performance of the three
techniques, namely, reconstruction based, interpolation
based and example based super resolution.
We used the flickr8k dataset across all three methods. All
images were cropped to dimensions of 300x300 and we
perform 4x upscaling of images. We measure the Peak
Signal to Noise Ratio (PSNR) in order to compare the
methods.
For interpolation techniques, we perform bilinear, bicubic,
area, nearest neighbour and Lanczos interpolation.
For a reconstruction based method using CNN to produce
low resolution images, we perform reconstruction of the
low resolution features to higher resolution images.
For example based methods we implemented a deep
convolutional network which uses skip connections to
preserve image quality. The PSNR as well as SSIM values
for these turned out to be better overall for this method
than for the previous methods. In terms of subjective
quality, the images generated by the deep network are
more visually pleasing than those generated by the other
methods.
5. RESULTS
Our experiments show that example based techniques
produce better images than interpolation and
reconstruction based methods. Example based techniques
also produce higher PSNR and SSIM scores than
interpolation and reconstruction based methods. Among
the different interpolation techniques, bicubic
interpolation consistently outperforms the other
interpolation techniques, producing a higher PSNR and
SSIM score.
Table1. PSNR and SSIM values comparison
Table 1. shows the PSNR and SSIM values for a particular
image of our professor that we obtained from the internet.
It can be observed that the PSNR values are higher for
Interpolation and Reconstruction based methods while
the SSIM values are better(higher the better) in Example-
based method.
6. CONCLUSIONS
Through our experiments we see that example based
techniques produce images that are of better quality than
images produced by interpolation and reconstruction
based techniques.
But even though our experimental results do produce
good quality images for experiment based techniques and
especially GANs, the PSNR values don’t seem to be the
right measure for proving the experimental results. While
the results obtained by deep learning are more pleasing to
the eye compared to other methods, they suffer from PSNR
statistics. We would like to analyse why deep learning
methods suffer from this and explore different loss
functions to make PSNR a suitable evaluation metric for
deep learning methods.
7. NEXT STEPS
The aim of our research project is to be able to generate
high resolution images using example based methods. By
implementing ideal GAN or CNN architectures using the
right activation and loss functions, our model should be
able to yield images of high quality and variety. We plan on
using an image Dataset like ImageNet and then train a GAN
model known to produce the best results. From the
literature survey, we can observe that PG-GAN is observed
to generate images of high quality and also follow a
continuous learning process and thus helps the
discriminator perform a better job at identifying new
samples generated by the generator. We plan on also
tweaking and adding changes to the basic GAN Loss
function by merging it with the Wasserstein Loss and the
Adam Optimizer. It has been established that PG-GAN does
not yet produce photorealistic images or generate micro-
structure details of images. Thus using PG-GAN along with
a strong loss function, our goal of research in improving
resolution of images could be realized hoping our model

succeeds in producing great results. The main goal is to
use a different accuracy measure other than the PSNR
value to determine the performance of example based
techniques and realize its power and conclude our model
as state of the art.
REFERENCES
[1] Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D.
Warde-Farley, S. Ozair, A. Courville, and Y. Bengio,
“Generative adversarial nets,” in Advances in Neural
Information Processing Systems, 2014, pp.2672–2680.
[2] E. L. Denton, S. Chintala, R. Fergus et al., “Deep
generative image models using a laplacian pyramid of
adversarial networks,”in Advances in Neural Information
Processing Systems, 2015, pp.1486–1494.
[3] A. Radford, L. Metz, and S. Chintala, “Unsupervised
representation learning with deep convolutional
generative adversarial networks,” in Proceedings of the
5th International Conference on Learning Representations
(ICLR) - workshop track, 2016.
[4] V. Dumoulin, I. Belghazi, B. Poole, O. Mastropietro, A.
Lamb, M. Arjovsky, and A. Courville, “Adversarially learned
inference,” in (accepted, to appear) Proceedings of the
International Conference on Learning Representations,
2017.
[5] J. Donahue, P. Krähenbühl, and T. Darrell, “Adversarial
feature learning,” in (accepted, to appear) Proceedings of
the International Conference on Learning Representations,
2017.
[6] K. Bousmalis, N. Silberman, D. Dohan, D. Erhan, and D.
Krishnan, “Unsupervised pixel-level domain adaptation
with generative adversarial networks,” in IEEE Conference
on Computer Vision and Pattern Recognition, 2016.
[7] S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and
H. Lee, “Generative adversarial text to image synthesis,” in
International Conference on Machine Learning, 2016.
[Online]. Available: https://arxiv.org/abs/1605.05396
[8] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired
image-to image translation using cycle-consistent
adversarial networks,” in Proceedings of the International
Conference on Computer Vision, 2017. [Online]. Available:
https://arxiv.org/abs/1703.10593
[9] C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Aitken, A.
Tejani, J. Totz, Z. Wang, and W. Shi, “Photo-realistic single
image super-resolution using a generative adversarial
network,” in IEEE Conference on Computer Vision and
Pattern Recognition, 2017.
[10] J. Liu, M. Spero, A. Raventos, “Super-Resolution on
Image and Video,”
[11] X. Xu, D. Sun, J. Pan, Y. Zhang, H. Pfister and M. Yang,
"Learning to Super-Resolve Blurry Face and Text Images,"
2017 IEEE International Conference on Computer Vision
(ICCV), Venice, 2017, pp. 251-260.
[12] T. Karras, T. Aila, S. Laine, J .Lehtinen (NVIDIA and
Aalto University), “Progressive Growing of GANs for
Improved Quality, Stability, and Variation,” in ICLR
Conference on Computer Graphics, Machine Learning and
Artificial Intelligence, 2018.
[13] Diederik P Kingma, Tim Salimans, Rafal Jozefowicz, Xi
Chen, Ilya Sutskever, and Max Welling. Improved
variational inference with inverse autoregressive flow. In
NIPS, volume 29, pp. 4743–4751. 2016.
[14] A. Beers, J. Brown, K. Chang, J. P. Campbell, S. Ostmo,
M. F. Chiang, and J. Kalpathy-Cramer. High resolution
medical image synthesis using progressively grown
generative adversarial networks. arXiv preprint
arXiv:1805.03144, 2018.
[15] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi
Huang. Gp-gan: Towards realistic high resolution image
blending. arXiv preprint arXiv:1703.07195, 2017.
[16] Curtó, J.D., Zarza, I.C., Torre, F.D.L., King, I. and Lyu,
M.R. [2018] High-resolution deep convolutional generative
adversarial networks. arXiv:1711.06491v9 [cs.CV].
[17] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-
image translation with conditional adversarial networks.
In CVPR, 2017
[18] O. Ronneberger, P. Fischer, and T. Brox. U-net:
Convolutional networks for biomedical image
segmentation. In MICCAI, pages 234–241. Springer, 2015
[19] T.D Harris, R.D Grober, J.K Trautman, Super
Resolution Image Spectroscopy, 1964
[20] Image Restoration: Fundamentals and Advance -
edited by Bahadir Kursat Gunturk, Xin Li
[21] L. Yue, H. Shen, J. Li, Q. Yuan, H. Zhang, and L. Zhang,
“Image super-resolution: The techniques, applications, and
future”
[22] Sabyasachi Moitra, “Single-Image Super-Resolution
Techniques: A Review”
[23] Aparna K V1 , Lisha P P2, “A Survey on Super-
Resolution Techniques”
[24] Kathiravan Srinivasan, “A Study on Super-Resolution
Image Reconstruction Techniques”
[25] Sina F, Dirk R, Michael E, Peyman M, “Advances and
Challenges in Super-Resolution”
[26] Lisha PP, Jayasree VK, “Single image super-resolution
- A quantitative comparison”

IRJET- Exploring Image Super Resolution Techniques

More Related Content

What's hot (20)

Similar to IRJET- Exploring Image Super Resolution Techniques (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Exploring Image Super Resolution Techniques