Image Super-Resolution Through Compressive Sensing-based Recovery
Hadi Zanddizari
Electrical Engineering
University of South Florida
Tampa, Florida, USA
hadiz@usf.edu
Ankita Dey
Systems and Computer Engineering
Carleton University
Ottawa, Canada
ankitadey@cmail.carleton.ca
Sreeraman Rajan
Systems and Computer Engineering
Carleton University
Ottawa, Canada
sreeraman.rajan@carleton.ca
Abstract— The primary aim of image super-resolution tech-
niques is to produce a high resolution (HR) image from a
low resolution (LR) image efficiently. Deep learning algorithms
are being extensively used to address the ill-posed problem of
single image super-resolution which requires extremely large
data-sets and high processing power. When one does not have
access to large data-sets or have limited processing power, an
alternative technique may be in order. In this study, we have
developed a novel positive scale image resizing method inspired
by compressive sensing (CS). We have considered the image
super-resolution as a CS recovery problem in which a low res-
olution image is assumed as a compressed measurement and the
required interpolated image is treated as output of the CS-based
recovery. In the proposed HR recovery method, a deterministic
binary block diagonal measurement matrix, (DBBD), is used
as measurement matrix since it maintains the visual similarity
between the low and high resolution images. Then along with a
sparsification matrix, the sparse representation of HR image
is first recovered and subsequently the dense HR image is
obtained. The proposed method is applied to medical and non-
medical images. The HR images obtained using the traditional
proximal, bilinear and bi-cubic interpolation techniques are
compared with those obtained using the proposed method. The
proposed CS inspired method delivers superior HR images
than the traditional techniques. The superiority of the proposed
method is attributed to the unique usage of the DBBD matrix
and the CS recovery algorithm to obtain a high resolution image
without any prior training data-set.
Index Terms–Compressive Sensing, Image interpolation, Im-
age Super-Resolution, Deterministic sensing matrix, Recovery
techniques
I. I NTRODUCTION
Images with high resolution help in making better di-
agnostic decisions from medical images. Often an image
with higher resolution (HR) will enable better detection of
anomalies such as tumors and cancerous cells than a low
resolution (LR) image. The quality and the resolution of an
image obtained through various medical imaging systems
such as X-rays, magnetic resonance imaging or computer
tomography play a crucial role in the diagnosis of a disease.
With high resolution images, it is possible to design auto-
matic diagnostic tools that could aid medical professionals
to make accurate decisions. It also enables doing object
detection and image segmentation with higher accuracy [3].
Realizing this, many diagnostic tools based on deep learning
models [1], [2] have been proposed lately. All these models,
however, require lot of images for training purposes. When
only one low resolution image is available, no learning
model can be developed. This paper aims to perform super-
resolution (SR) from just a single low resolution (LR) image
under the assumption that there is no access to any huge
dataset.
The term SR can be defined as a technique to enhance
or increase the resolution of an image. Single image super-
resolution (SISR) aims to generate a HR image from a LR
one. Traditionally, super-resolution is attempted either using
multiple images and solving for a set of linear constraints or
by learning relationship between LR and HR image patches
from a database (called example-based approach) [4]. Since
the mapping between LR image and HR images are not
unique, SISR is an ill-posed problem for image recovery
[5]. Although multiple HR images were used in [6] to
reconstruct the HR image, the reconstructed image was not
guaranteed to contain true HR details. Since recovery of HR
was based on multiple examples, new learning algorithms
such Bayesian approach [7], neighbour embedding method
[8], recovery using sparse patches from LR images [9]
were introduced. Over the years, interestingly even for SISR
problem, deep learning (DL)-based approaches have become
the most sought method.
Amongst the DL-based approaches, the SR convolu-
tional neural network (CNN) has become the benchmark
architecture for DL-based SR algorithm [10]. Deep neural
network based unsupervised algorithms such as the deep
Boltzmann machine [11], variational autoencoder (VAE) [12]
and generative adversarial nets (GAN) [13] have also been
implemented to handle unlabeled data situations. All these
methods claim their superiority in terms of accuracy but
do not highlight the shortcomings associated with them.
Firstly, DL algorithms need large pre-trained data-sets for
efficient mapping (or learning) which may not be available
in many problems. Secondly, DL techniques are image
specific. Thirdly, the DL networks may have the problem of
over-fitting. Particularly for medical images, such erroneous
results may lead to wrong diagnosis. Lastly, implementation
of DL algorithms are computationally intensive and therefore
require computers with huge processing capability. Unfortu-
nately, DL methods are of no avail when only one LR image
is available and needs to be converted into a SR image.
To overcome the aforementioned shortcomings, non-
learning based algorithms may be used. Traditional predic-
tion models or interpolation techniques such as proximal,
bilinear or bicubic interpolation use weighted average neigh-
bouring LR pixel intensities to generate a HR image. The
2021 43rd Annual International Conference of the
IEEE Engineering in Medicine & Biology Society (EMBC)
Oct 31 - Nov 4, 2021. Virtual Conference
978-1-7281-1178-0/21/$31.00 ©2021 IEEE 4006