Image Super Determination Model Enabled through Wavelet Lifting with Improved Deep Convolutional Neural Network

Abstract

Super resolution techniques have emerged for different applications in various fields as HR
photographs provide an abundance of information and helpful. Deep learning-based SR models have been
rapidly developed over the past few years, and deep learning-based SR models are commonly found to
deliver on the benchmarks of super resolution photos. This study proposes to use Wavelet-based highresolution
processing with Deep learning algorithms for super-resolution processing. The resolution of the
HR images is changed to Low Resolution (LR) utilizing bicubic interpolation-based down sampling and up
sampling before beginning the resolution step. It is also helpful to generate the four sub-bands of each image
by employing the Wavelet lifting approach. Also, the residual picture is created by subtracting the LRSB
from the HRSB. Training and testing are the two major phases in the suggested paradigm. During the
training phase, Deep Convolutional Neural Network (DCNN) trains the residual image of all images by
feeding LRSB as input and the residual image as target. This alternative view is in line with the idea that in
testing, the LRSB query image is subject to Deep CNN, which returns the residual image to the neural
network. The image that was created as a result of the summing of the residual image with LRSB image,
along with an inverse wavelet lifting procedure, results in the final super resolution image. This work aims
to improve the Deep CNN by modifying the Whale Optimization Algorithm (WOA) by altering the number
of hidden layers and hidden neurons (PSNR). In the end, the proposed approach delivered equivalent
outcomes to the other models.