Loading [MathJax]/extensions/TeX/ietmacros.js
Improving resolution of images using Generative Adversarial Networks | IEEE Conference Publication | IEEE Xplore

Improving resolution of images using Generative Adversarial Networks


Abstract:

Even with all the achievements in precision and speed of various image super-resolution models, such as better and more accurate Convolutional Neural Networks (CNN), the ...Show More

Abstract:

Even with all the achievements in precision and speed of various image super-resolution models, such as better and more accurate Convolutional Neural Networks (CNN), the results have not been satisfactory. The high-resolution images produced are generally missing the finer and frequent texture details. The majority of the models in this area focus on such objective functions which minimize the Mean Square Error (MSE). Although, this produces images with better Peak Signal to Noise Ratio (PSNR) such images are perceptually unsatisfying and lack the fidelity and high-frequency details when seen at a high-resolution. Generative Adversarial Networks (GAN), a deep learningmodel, can be usedfor such problems. In this article, the working of the GAN is shown and described about the production satisfying images with decent PSNR score as well as good Perceptual Index (P1) when compared to other models. In contrast to the existing Super Resolution GAN model, various modifications have been introduced to improve the quality of images, like replacing batch normalization layer with weight normalization layer, modified the dense residual block, taking features for comparison before they are fed in activation layer, using the concept of a relativistic discriminator instead of a normal discriminator that is used in vanilla GAN and finally, using Mean Absolute Error in the model.
Date of Conference: 05-07 November 2020
Date Added to IEEE Xplore: 28 December 2020
ISBN Information:
Conference Location: Coimbatore, India

I. Introduction

Image super-resolution basically means improving or enhancing the resolution of an image. It is a problem to convert low-resolution image to high-resolution image or simply, to recover the image’s finer texture details. It has got attention over the time from many research communities as it has various practical applications [11, 12, 13]. There have been many Super Resolution (SR) algorithms such as SRCNN [43], introduced by Dong et al. [4, 5] which focus on the minimization of Mean Squared Error and thus, in tum, maximizes the Peak Signal to Noise Ratio [7], which is a wellknown metric for comparing the quality of images [8]. Though there have been many breakthroughs in this field with such models, and performance has been improved over time but such algorithms tend to produce results that lack perceptual clarity and fine texture details [44]. This is because MSE and PSNR aren’t able to do justification with the perceptually important features of an image as these metrics focus on the differences in pixels of the recovered image and ground truth image (generated super-resolved image and original highresolution image). The resultant images look so unrealistically smooth and devoid of any high-frequency detail that it fails in the subjective evaluation [2].

References

References is not available for this document.