Super Resolution (SR) Generative Adversarial Networks (GANs) for Feature Extraction on Different Datasets

Abstract

There have been tremendous advancements in using Generative Adversarial Network (GAN) models for computer vision. Despite new advancements in GANs for computer vision, super-resolution (SR) is still considered a challenging research topic in computer vision. Super-resolution has various challenges such as ill-posed inverse problem, complexity growth with increase in up-scaling factor and complexity in assessment of output quality. In recent years, there is a surge of interest in super-resolution methods using Generative Adversarial Networks to solve these challenges. Existing research has been mostly focused on few key techniques and single dataset for feature extraction and calculations. So far, there hasn’t been comprehensive research to study various GANs models for existing State-of-the-Art (SOTA) in super-resolution methods for feature extraction on multiple datasets.
The purpose of this research is to study and improve SOTA in super-resolution GANs through extending and modifying various existing architectures and try new architectures for feature extraction to evaluate results on multiple datasets. This paper focuses on VGG19, VGGFace2 and EfficientNet as pre-trained transfer learning backbones for feature extraction within super-resolution GANs on multiple large challenging face images datasets (CelebA, LFW and Simpson). The result will be evaluated using Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics.