Smart and Sustainable Intelligent Systems

Namita Gupta, Prasenjit Chatterjee, Tanupriya Choudhury

• Publisher: Wiley-Scrivener
• Year: 2021
• ISBN:
  9781119752103
• Edition: 1
• Topic: Computer Science
• Subtopic: Artificial Intelligence (AI) & Semantics
• Index: Computer Science

The world is experiencing an unprecedented period of change and growth through all the electronic and technilogical developments and everyone on the planet has been impacted. What was once 'science fiction', today it is a reality.

This book explores the world of many of once unthinkable advancements by explaining current technologies in great detail. Each chapter focuses on a different aspect - Machine Vision, Pattern Analysis and Image Processing - Advanced Trends in Computational Intelligence and Data Analytics - Futuristic Communication Technologies - Disruptive Technologies for Future Sustainability. The chapters include the list of topics that spans all the areas of smart intelligent systems and computing such as: Data Mining with Soft Computing, Evolutionary Computing, Quantum Computing, Expert Systems, Next Generation Communication, Blockchain and Trust Management, Intelligent Biometrics, Multi-Valued Logical Systems, Cloud Computing and security etc. An extensive list of bibliographic references at the end of each chapter guides the reader to probe further into application area of interest to him/her.

Part 1
MACHINE LEARNING AND ITS APPLICATION

1
Single Image Super-Resolution Using GANs for High-Upscaling Factors

Harshit Singhal*, Aman Kumar, Shubham Khandelwal, Anupam Kumar and Mini Agarwal

Maharaja Agrasen Institute Of Technology, Guru Gobind Singh Indraprastha University, Dwarka, New Delhi, India

Abstract

Some state-of-the-art networks have solved the problem of recovering photorealistic features up to a certain level of up-sampling factor, basically up-to a factor of 4× in the task of Single Image Super-Resolution, but recovering fine and photorealistic features in up-sampling factors of 8× and above still remains unsolved. To solve this problem, we studied some components of SRGAN and came up with some improvements in these networks. Firstly, we propose to improve the residual block architecture to define a modified residual dense block with the addition of the batch-normalization layer and secondly, we improvised the Perceptual loss function according to our GAN. After these two changes, we introduce a new GAN-based architecture for the Single Image Super-Resolution task for higher up-sampling levels.

Keywords: Super resolution, 8× upsampling, GANs, higher upsampling, higher upscaling, dense block for upsampling, realistic super resolution, single image upsampling

1.1 Introduction

Image Super Resolution is one of the problems in Computer Science which has numerous applications. Single Image Super Resolution is the problem which is defined as converting a Low-Resolution Image into a High-Resolution Image without losing the Quality of the image and still capturing fine photo-realistic textures. The applications of this class of problem are in the field of medical science [1], in Digital Image Processing [2] and in various other fields [3, 4] are of great importance.

A very practical example of Image Super resolution is what we see in movies and shows where someone zooms into a picture and the quality improves and detail appears as shown in Figure 1.1. A very interesting use case of this is imagine transferring a picture on a network where you are sending 128 × 128 resolution image where 1,024 × 1,024 image is needed. It would also be beneficial in many other areas e.g., medical applications, aerospace, etc.

Figure 1.1 Low resolution–high resolution.

The approach we are trying is to solve this problem using GANs. Since traditional algorithm-based upscaling methods usually do not capture the finer details, humanly, this task is lengthy and too much time-consuming.

Obviously, on applying a degradation method available in the industry, we can acquire the Low-Resolution Image from the High-Resolution Image. Yet, would we be able to do the reverse? Of course, we can in an ideal situation if we know the method which was used to degrade the image in the first place, by applying its opposite to the LR picture, we can regenerate the HR picture. Problem in most cases is that we are just provided the Low-Resolution image and we are expected to upgrade it without knowing the source of the image, since we do not know the source of image, we can’t find the degradation method used and can’t recover the High-resolution image.

The task of single image super resolution is also a very ill-posed problem due to multiple possible outputs for the same low-resolution image. Since in low resolution images a lot of information is lost, for example if we have a low-resolution image which needs to be converted to 8× up sampling factor, then a single pixel in low resolution image corresponds to 16 pixels in the higher resolution image. So basically a 64-pixel information is compressed and degraded to a single pixel. This information needs to be generated for which various methods including interpolation techniques [5, 6] were used. But they did not result in regenerating the higher resolution image in desired manner which led to the use of Deep Learning based methodologies for solving this problem. Good results were obtained when Convolutional Neural Networks were used for this task [7, 8], but the use of Convolutional Neural Networks were not able to regenerate the fine textural details. This is where supervised learning approach were unable to properly solve this problem. Then the unsupervised learning based approach like GANs came into picture with the evolution of SRGAN [9] which was among the first GANs-based super resolution solution.

With detailed study of SRGAN, our work reflects how the basic structure of SRGAN can be modified to get better performance in the task of super resolution. Here in this work we will focus on how Residual Block architecture in SRGAN is modified and in accordance to it how generator is modified to get better results for 8× up sampling factors.

1.2 Methodology

1.2.1 Architecture Details

1. 1. Generator Architecture: The task of Single Image Super Resolution involves providing only a Single Low-Resolution Image (denoted by I_LR) and generating a High-Resolution image output (denoted as I_OHR). In the following work, we describe a GANs generator network which helps to obtain photo-realistic features in I_OHR which is very identical to the Ground Truth High-Resolution Image (denoted as I_GHR).
   The previous work on super resolution task involves SRGAN [9], which has a very good Generator and Discriminator network architecture. Some works on SRGAN modified the Residual Block Architecture to get better building block for the Generator of SRGAN like in Ref. [10].
   The architecture shown in Figure 1.2 is a representation of how I_LR can be converted to High Resolution Image by a factor of 8×.
   The network consists of a Input layer which ...

Table of contents

• Cover
• Table of Contents
• Title Page
• Copyright
• Dedication
• Preface
• Acknowledgement
• Part 1: MACHINE LEARNING AND ITS APPLICATION
• Part 2: DEEP LEARNING AND ITS APPLICATION
• Part 3: SECURITY AND BLOCKCHAIN
• Part 4: COMMUNICATION AND NETWORKS
• Part 5: LATEST TRENDS IN SUSTAINABLE COMPUTING TECHNIQUES
• Index
• End User License Agreement