Experimentally Proven Bilateral Blur on SRCNN for Optimal Convergence

DR. DIVYA MISHRA, Tayyibah Khanam, Ivanshu Kaushik

Abstract:

Image super-resolution is a contentious issue in developing
computer vision applications such as satellite imaging, graphics indus-
try, medical diagnostics, and real-time scene monitoring for security and
safety. CNN-based image super-resolution algorithms are the simplest
and most resource-efficient in deep learning. The original SRCNN frame-
work, on the other hand, incorporated Gaussian blur, which takes a long
time to converge. This is not only computationally expensive, but it also
extends the training time. In light of this, we retrained original SRCNN
utilizing widely available blurring techniques and observed that the bilat-
eral filter beats the Gaussian filter at optimal convergence. Our goal is
to preserve the original SRCNN features while reducing training time
and computational resources and hence, speed up the architecture with
optimum number of epochs. This is, to the best of our knowledge, the
first study of its kind, and no other study has implemented this alterna-
tive blurring training on SRCNN and ranked the best of them for image
super-resolution.