Image super resolution using an improved generative adversarial network framework with image enlargement algorithms and convolutional neural network

Abstract

High-resolution images are essential in various fields, such as medical imaging, newlinesatellite imaging, security and surveillance, entertainment, forensics, self-driving cars, newlinemicroscopy, etc. Humans or machines can better identify and analyze objects with newlineimproved image resolution. Although hardware devices can capture high-quality newlineimages, they can be expensive and require maintenance. On the other hand, lowresolution newlinedevices are affordable but produce images that are difficult to identify. newlineSuper-resolution is a technique used in image processing to convert low-resolution newlineimages into high-resolution images with improved quality. Traditional methods were newlineused to produce high-resolution images, but deep-learning methods have made newlinesignificant progress in image processing in recent years. Deep learning has resulted in newlinebetter performance in image super-resolution than traditional methods. However, newlinethere is still a gap between the results of current practices and the desired highresolution newlinequality in real-world scenarios. newlineThe proposed research presents a comprehensive framework for super-resolution newlineimage generation using a dual-generator Generative Adversarial Network (GAN). The newlineframework combines pre-processing techniques, GAN architecture with Squeeze and newlineExcitation blocks, and image fusion to achieve superior image resolution and quality. newlineThe pre-processing step is crucial in enlarging the input image using enlarging newlinealgorithms. This prepares the image for the subsequent enhancement process, newlineallowing the GAN to focus solely on improving the image details and quality. newlineEnlarging algorithms at different scales (2x and 4x) provides flexibility in catering to newlinedifferent resolution requirements. The GAN architecture incorporates dual generators, newlinewhich work collaboratively to generate high-resolution images. Including Squeeze newlineand Excitation blocks enhances the channel interdependencies within the network, newlineresulting in improved efficiency and performance. The proposed framework newlinesignificantly improves super-resolution image gene