Image Super Resolution and Deblurring using Generative Adversarial Networks

As deep learning methodologies exhibit strong advantages in the vast domain of feature extraction, it has been used widely in the computer vision area, and gradually replaced traditional machine learning algorithms. This paper describes in detail the development, creation and Analytics of a Deep learning pipeline which is comprised of two components firstly a Deblur Generative Adversarial Network (DeblurGAN) and secondly a Super Resolution Generative Adversarial Net-work (SRGAN) which work in unison with one another in order to convert and transform Low Quality Blurred images into a higher quality Super resolved deblurred image. In our solution we are aiming to tackle various types of blurs such as linear blurring, Gaussian Blurring and Media Blurring. Generative adversarial networks (GANs) are algorithmic architectures that primarily use two neural networks, pitting one against the other (hence the ”adversarial”) in order to produce new ,enhanced, synthetic data instances that can pass for real data, using this principle we are generating de-blurred and super-resolved images. The Trained Discriminator acts as an accurate classifier and has the main functionality of differentiating and distinguishing between the multiple generated images of the GAN and the real images. At the specific point in time when the discriminator component is unable to differentiate between the images that are created and the actual images, it is safe to say that we have generated the most de-blurred and Super-resolved image possible. Afterwards the results (i.e. Peak Signal to Noise Ratio aka PSNR) are compared in different scenarios to observe which type of blur is best tackled by the deep learning pipeline.