End-to-End Image Super-Resolution via Deep and Shallow Convolutional Networks

Abstract

A novel picture super-resolution (SR) technique based on a
Convolution Neural Network (CNN) is being developed as part of
this project's research. When learning the feature extraction,
upsampling, and high-resolution (HR) reconstruction modules at
the same time, a deep convolutional neural network (CNN) is
created that can be used to rebuild pictures from any source and
is completely trainable. If, on the other hand, you want to train a
deep network in a straight line from start to end, this is timeconsuming
and may provide sub-optimal results since it ta k es a
longer time to converge than other strategies. According to our
results, an ensemble of deep and shallow networks should be
trained at the same time in order to overcome this difficul ty. Its
stronger representation power, rather than a lower learning
capacity, allows the deep network to capture the high-frequency
information contained within visual images, rather than the
other way around. When utilised in combination with joint
training, the shallow network reduces the complexity of deep
network optimization by a factor of two, in part because the
shallow network is considerably simpler to optimise than the deep
network. High frequency characteristics are rebuilt in a multiscale
manner to further improve the accuracy of HR
reconstruction. This allows for the simultaneous integration of
both short- and long-range contextual information to be included
in the reconstruction, which further improves the accuracy of
HR reconstruction. The suggested technique has been careful ly
examined on a variety of commonly used data sets, and when
compared to current best practises, it beats them by a significant
margin. Large-scale ablation experiments are carried out to
establish the contributions of various network topologies to
image SR, which results in the finding of new insights that may
be used to future study.