Generative Adversarial Network for Super

Resolution and Blind Image Quality

*Note: Sub-titles are not captured in Xplore and should not be used

Ammara Nasim Osama Amjad

Department of Electrical Engineering, Department of Electrical Engineering,
Bahria University Bahria University
Islamabad, Pakistan Islamabad, Pakistan
ammaranaseem.buic@bahria.edu.pk osamamjad26@gmail.com

Ammar Ahmed Bushra Amjad

Department of Electrical Engineering, Department of Computer Science,
Bahria University National University of Computer and Emerging Sciences, FAST
Islamabad, Pakistan Islamabad, Pakistan
ammar22041@gmail.com bushramjad16@gmail.com

Abstract—This electronic document is a “live” template and We have used a deep residual network (ResNet) with
already defines the components of your paper [title, text, heads, skip-connection and diverge from MSE as the sole
etc.] in its style sheet. *CRITICAL: Do Not Use Symbols, Special optimization target. They developed a novel perceptual loss
Characters, Footnotes, or Math in Paper Title or Abstract. by combining high-level feature maps from the
(Abstract) VGG network with a discriminator that encourages solutions
that are perceptually difficult to distinguish from HR
Keywords—component, formatting, style, styling, insert (key reference images. A photorealistic image that was super-
words)
resolved with a 4x up-scaling factor and for quality
I. INTRODUCTION assessment we have used Type Classification and Label
Transfer (TCLT) [3].
Super-resolution imaging (SR) refers to a group of
techniques that improve (increase) an imaging system's Image quality measurements are crucial in a wide range
resolution. The particularly complex task of estimating a of image processing applications. The aim of image quality
high-resolution (HR) image from its low-resolution (LR) assessment (IQA) is to provide a quality metric that can be
counterpart is known as super-resolution (SR). The used to determine the efficiency of image processing systems.
diffraction limit of systems is overcome in optical SR, The IQA offers computational models for assessing an
whereas the resolution of digital imaging sensors is improved image's perceptual performance. These methods are typically
in geometrical SR because a specific low-resolution (LR) divided into three categories based on the availability of a
input can correspond to a crop of possible high-resolution reference image: full-reference (FR), reduced-reference
(HR) images, and the HR space that we intend to map the LR (RR), and no-reference/blind (NR). The NR/Blind-IQA
input to is usually intractable, single image super-resolution approach does not need the reference image details, which is
(SISR) is a notoriously difficult ill-posed problem. The both appealing and challenging.
disadvantages of previous SISR approaches are primarily In this BIQA, a label transfer method as proposed by Q.
two. Previous SISR methods have primarily two drawbacks: Wu et al. [3] to intuitively simulate the visual memory
the first is an unclear definition of the mapping that we hope retrieval process. The visual perception properties underlying
to develop between the LR space and the HR space, and the each proposed feature are investigated. To evaluate the
second is the inefficiency of establishing a complex high proposed method, more extensive experiments and
dimensional mapping given massive raw data. Taking application examples are added. In the meantime, the
advantage of a strong ability to extract effective high-level computational complexity of our method is being
abstractions that bridge the LR and HR space. investigated.
The texture information is typically missing in The majority of existing Blind Image Quality Assessment
reconstructed SR images, the underdetermined SR problem is (BIQA) approaches are made up of two modules: 1) Quality-
caused for high up-scaling factors. Supervised SR algorithms aware feature extraction. This module generates an efficient
normally aim to reduce the mean squared error (MSE) image depiction to capture the perceptual quality variation
between the recovered HR image and the ground reality. This caused by distortion. As discussed in [2, 14–17], many BIQA
is convenient because minimizing MSE also maximizes the methods focus on describing an image based on its natural
peak signal-to-noise ratio (PSNR), which is a common metric scene statistics (NSS) from a single color channel. 2)
used to evaluate and compare SR algorithms. Also, digital Prediction model learning. This module's primary function is
images are widely used in visual communications, to map image features to subjective quality scores. Learning-
entertainment, and social networking sites. In these fields, an based regression models, such as support vector regression
effective image quality assessment (IQA) algorithm [1] is (SVR) [21] and general regression neural network (GRNN)
critical for evaluating, controlling, and improving perceptual [22], are widely used in many BIQA algorithms [15], [16],
image quality. [1] [18]–[20].
In this paper, we used a super-resolution generative
adversarial network (SRGAN) proposed by Ledig et al. [2].

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

 II. RELATED WORK Lim et al. [52] propose an EDSR model by removing
Recently published overview articles on image SR unnecessary BN layers in the residual block and increasing
include prediction-based methods, which were among the model size. Zhang et al. [53] propose using an effective
first to address SISR. While these filtering approaches, such residual dense block in SR, and they go on to investigate a
as linear, bicubic, or Lanczos [1] filtering, can be very fast, deeper network with channel attention [54]. Other methods,
they oversimplify the SISR problem and typically produce such as reinforcement learning [55] and unsupervised
solutions with excessively smooth textures. Methods with a learning [56], are introduced in addition to supervised
special emphasis on edge preservation have been proposed learning to solve general image restoration problems.
[2, 39]. More powerful approaches, which typically rely on Several methods for stabilizing training in a very deep
training data, aim to establish a complex mapping between model have been proposed.
low- and high-resolution image information. Many example
pair-based methods rely on LR training patches for which the A residual path, for example, is created to stabilize training
corresponding HR counterparts are known. Freeman et al. and improve performance [57, 58, 54]. Residual scaling was
[19, 18] presented early work. Compressed sensing is the pioneered by Szegedy et al. [59] and is now widely used in
origin of related approaches to the SR problem [40, 13, 41]. EDSR. He et al. [60] propose a robust initialization method
Glasner et al. [22] use patch redundancy across scales within for VGG-style networks without BN for general deep
the image to drive the SR. This self-similarity paradigm is networks.
also used by Huang et al. [32], who extend self-dictionaries Photo-realism is typically achieved through adversarial
by allowing for small transformations and shape variations. training with GAN [61]. There have recently been several
Gu et al. [26] proposed a convolutional sparse coding method works focusing on developing more effective GAN
that improves consistency by processing the entire image frameworks. WGAN [62] proposes minimizing a reasonable
instead of overlapping patches. and efficient approximation of the Wasserstein distance and
Kim et al. [35] demonstrated a highly performant regularizing the discriminator through weight clipping.
architecture that allows for long-range pixel dependencies Gradient clipping is another improved regularization for
while keeping the number of model parameters small using a discriminator [63], and spectral normalization [64].
deeply recursive convolutional network (DRCN). Johnson et SR algorithms are typically assessed using a variety of
al. [34] and Bruna et al. [6] recover visually convincing HR widely used distortion measures, such as PSNR and SSIM.
images by using a loss function that is closer to perceptual These metrics, however, fundamentally disagree with human
similarity. observers' subjective evaluations [46]. Non-reference
Following the success of Krizhevsky et al. [38], the state measures, such as Ma's score [65] and NIQE [66], are used to
of the art for many computer vision problems is now set by assess perceptual quality in the PIRM-SR Challenge [67].
specifically designed CNN architectures. Following the III. PREPARE YOUR PAPER BEFORE STYLING
success of Krizhevsky et al. [38], the state of the art for many
computer vision problems is now set by specifically designed Before you begin to format your paper, first write and save
CNN architectures. Deeper network architectures have also the content as a separate text file. Complete all content and
been shown to improve SISR performance; for example, Kim organizational editing before formatting. Please note sections
et al. [35] present state-of-the-art results for a recursive CNN. A-D below for more information on proofreading, spelling
The recently introduced concept of residual blocks [30] and and grammar.
skip connections [31, 35] is another powerful design choice Keep your text and graphic files separate until after the text
that simplifies deep CNN training. has been formatted and styled. Do not use hard tabs, and limit
Glasner et al. [42] proposed a unified framework for use of hard returns to only one return at the end of a paragraph.
integrating two methods. The traditional multi-image super- Do not add any kind of pagination anywhere in the paper. Do
resolution (combining images obtained at sub pixel not number text heads-the template will do that for you.
misalignments), and (ii) Example-Based super-resolution. A. Abbreviations and Acronyms
They demonstrated how to use this combined approach to
obtain super-resolution from a single image (with no database Define abbreviations and acronyms the first time they are
or prior examples). Their method attempts to recover the best used in the text, even after they have been defined in the
possible resolution increase at each pixel based on patch abstract. Abbreviations such as IEEE, SI, MKS, CGS, sc, dc,
redundancy within and across scales. Experiments show that and rms do not have to be defined. Do not use abbreviations
the Example-Based SR component in our combined in the title or heads unless they are unavoidable.
framework provides the greatest improvement in resolution. B. Units
The Classical-SR component, on the other hand, is critical in
preventing the Example-Based SR component from  Use either SI (MKS) or CGS as primary units. (SI units
hallucinating incorrectly. are encouraged.) English units may be used as
secondary units (in parentheses). An exception would
SRCNN was proposed by Dong et al. [43,44] to learn the be the use of English units as identifiers in trade, such
mapping from LR to HR images in 4 Xintao Wang et al. in as “3.5-inch disk drive”.
an end-to-end manner, outperforming previous works. Later,
the field saw a variety of network architectures, including a  Avoid combining SI and CGS units, such as current in
deeper network with residual learning [21], Laplacian amperes and magnetic field in oersteds. This often
pyramid structure [45], residual blocks [46], and recursive leads to confusion because equations do not balance
learning [47,48]. Densely connected network [49], deep back dimensionally. If you must use mixed units, clearly
projection [50], and residual dense network [51]. state the units for each quantity that you use in an
equation.
  Do not mix complete spellings and abbreviations of “compliment”, “discreet” and “discrete”, “principal”
units: “Wb/m2” or “webers per square meter”, not and “principle”.
“webers/m2”. Spell out units when they appear in text:
“. . . a few henries”, not “. . . a few H”.  Do not confuse “imply” and “infer”.

 Use a zero before decimal points: “0.25”, not “.25”.  The prefix “non” is not a word; it should be joined to
Use “cm3”, not “cc”. (bullet list) the word it modifies, usually without a hyphen.

C. Equations  There is no period after the “et” in the Latin

abbreviation “et al.”.
The equations are an exception to the prescribed
specifications of this template. You will need to determine  The abbreviation “i.e.” means “that is”, and the
whether or not your equation should be typed using either the abbreviation “e.g.” means “for example”.
Times New Roman or the Symbol font (please no other font). An excellent style manual for science writers is [7].
To create multileveled equations, it may be necessary to treat
the equation as a graphic and insert it into the text after your IV. USING THE TEMPLATE
paper is styled.
After the text edit has been completed, the paper is ready
Number equations consecutively. Equation numbers, for the template. Duplicate the template file by using the Save
within parentheses, are to position flush right, as in (1), using As command, and use the naming convention prescribed by
a right tab stop. To make your equations more compact, you your conference for the name of your paper. In this newly
may use the solidus ( / ), the exp function, or appropriate created file, highlight all of the contents and import your
exponents. Italicize Roman symbols for quantities and prepared text file. You are now ready to style your paper; use
variables, but not Greek symbols. Use a long dash rather than the scroll down window on the left of the MS Word
a hyphen for a minus sign. Punctuate equations with commas Formatting toolbar.
or periods when they are part of a sentence, as in:
A. Authors and Affiliations
 ab  The template is designed for, but not limited to, six
authors. A minimum of one author is required for all
Note that the equation is centered using a center tab stop. conference articles. Author names should be listed starting
Be sure that the symbols in your equation have been defined from left to right and then moving down to the next line. This
before or immediately following the equation. Use “(1)”, not is the author sequence that will be used in future citations and
“Eq. (1)” or “equation (1)”, except at the beginning of a by indexing services. Names should not be listed in columns
sentence: “Equation (1) is . . .” nor group by affiliation. Please keep your affiliations as
succinct as possible (for example, do not differentiate among
D. Some Common Mistakes departments of the same organization).
 The word “data” is plural, not singular. 1) For papers with more than six authors: Add author
 The subscript for the permeability of vacuum 0, and names horizontally, moving to a third row if needed for more
other common scientific constants, is zero with than 8 authors.
subscript formatting, not a lowercase letter “o”. 2) For papers with less than six authors: To change the
 In American English, commas, semicolons, periods, default, adjust the template as follows.
question and exclamation marks are located within a) Selection: Highlight all author and affiliation lines.
quotation marks only when a complete thought or b) Change number of columns: Select the Columns
name is cited, such as a title or full quotation. When icon from the MS Word Standard toolbar and then select the
quotation marks are used, instead of a bold or italic correct number of columns from the selection palette.
typeface, to highlight a word or phrase, punctuation c) Deletion: Delete the author and affiliation lines for
should appear outside of the quotation marks. A the extra authors.
parenthetical phrase or statement at the end of a
sentence is punctuated outside of the closing
parenthesis (like this). (A parenthetical sentence is B. Identify the Headings
punctuated within the parentheses.)
Headings, or heads, are organizational devices that guide
 A graph within a graph is an “inset”, not an “insert”. the reader through your paper. There are two types:
The word alternatively is preferred to the word component heads and text heads.
“alternately” (unless you really mean something that
Component heads identify the different components of
alternates).
your paper and are not topically subordinate to each other.
 Do not use the word “essentially” to mean Examples include Acknowledgments and References and, for
“approximately” or “effectively”. these, the correct style to use is “Heading 5”. Use “figure
caption” for your Figure captions, and “table head” for your
 In your paper title, if the words “that uses” can table title. Run-in heads, such as “Abstract”, will require you
accurately replace the word “using”, capitalize the “u”; to apply a style (in this case, italic) in addition to the style
if not, keep using lower-cased. provided by the drop down menu to differentiate the head
 Be aware of the different meanings of the homophones from the text.
“affect” and “effect”, “complement” and Text heads organize the topics on a relational, hierarchical
Identify applicable funding agency here. If none, delete this text box. basis. For example, the paper title is the primary text head
because all subsequent material relates and elaborates on this REFERENCES
one topic. If there are two or more sub-topics, the next level The template will number citations consecutively within
head (uppercase Roman numerals) should be used and, brackets [1]. The sentence punctuation follows the bracket [2].
conversely, if there are not at least two sub-topics, then no Refer simply to the reference number, as in [3]—do not use
subheads should be introduced. Styles named “Heading 1”, “Ref. [3]” or “reference [3]” except at the beginning of a
“Heading 2”, “Heading 3”, and “Heading 4” are prescribed. sentence: “Reference [3] was the first ...”
C. Figures and Tables Number footnotes separately in superscripts. Place the
a) Positioning Figures and Tables: Place figures and actual footnote at the bottom of the column in which it was
tables at the top and bottom of columns. Avoid placing them cited. Do not put footnotes in the abstract or reference list. Use
in the middle of columns. Large figures and tables may span letters for table footnotes.
across both columns. Figure captions should be below the Unless there are six authors or more give all authors’
figures; table heads should appear above the tables. Insert names; do not use “et al.”. Papers that have not been
figures and tables after they are cited in the text. Use the published, even if they have been submitted for publication,
abbreviation “Fig. 1”, even at the beginning of a sentence. should be cited as “unpublished” [4]. Papers that have been
accepted for publication should be cited as “in press” [5].
TABLE I. TABLE TYPE STYLES
Capitalize only the first word in a paper title, except for proper
Table Table Column Head nouns and element symbols.
Head Table column subhead Subhead Subhead For papers published in translation journals, please give
copy More table copy a
the English citation first, followed by the original foreign-
a. language citation [6].
Sample of a Table footnote. (Table footnote)

Fig. 1. Example of a figure caption. (figure caption)

[1] C. E. Duchon. Lanczos Filtering in One and Two Dimensions. In
Figure Labels: Use 8 point Times New Roman for Figure Journal of Applied Meteorology, volume 18, pages 1016–1022. 1979
labels. Use words rather than symbols or abbreviations when [2] C. Ledig et al., “Photo-realistic single image super-resolution using a
generative adversarial network,” Proc. - 30th IEEE Conf. Comput. Vis.
writing Figure axis labels to avoid confusing the reader. As an Pattern Recognition, CVPR 2017, vol. 2017-Janua, pp. 105–114, 2017,
example, write the quantity “Magnetization”, or doi: 10.1109/CVPR.2017.19
“Magnetization, M”, not just “M”. If including units in the [3] Q. Wu et al., “Blind image quality assessment based on multichannel
label, present them within parentheses. Do not label axes only feature fusion and label transfer,” IEEE Trans. Circuits Syst. Video
with units. In the example, write “Magnetization (A/m)” or Technol., vol. 26, no. 3, pp. 425–440, 2016, doi:
“Magnetization {A[m(1)]}”, not just “A/m”. Do not label axes 10.1109/TCSVT.2015.2412773.
with a ratio of quantities and units. For example, write [4]
“Temperature (K)”, not “Temperature/K”. IEEE conference templates contain guidance text for
composing and formatting conference papers. Please
ACKNOWLEDGMENT (Heading 5) ensure that all template text is removed from your
The preferred spelling of the word “acknowledgment” in conference paper prior to submission to the
America is without an “e” after the “g”. Avoid the stilted conference. Failure to remove template text from
expression “one of us (R. B. G.) thanks ...”. Instead, try “R. your paper may result in your paper not being
B. G. thanks...”. Put sponsor acknowledgments in the published.
unnumbered footnote on the first page.

We suggest that you use a text box to insert a graphic

(which is ideally a 300 dpi TIFF or EPS file, with all fonts
embedded) because, in an MSW document, this method is
somewhat more stable than directly inserting a picture.
To have non-visible rules on your frame, use the
MSWord “Format” pull-down menu, select Text Box >
Colors and Lines to choose No Fill and No Line.