Image Super-Resolution with GANs

Problem Statement

Image Super-Resolution with GANs: Train a GAN to generate high-resolution images from low-resolution inputs. 

Introduction 

•With the advancement of deep learning, Generative Adversarial Networks (GANs) have emerged as powerful tools for generating high-quality images.

•The adversarial framework helps the generator learn to produce perceptually convincing images, while the discriminator pushes it to generate outputs that are indistinguishable from real high-resolution images.

•The model learns how to "fill in" the missing details in a smart and realistic way by training on a large set of example images. After training, it can take a small or blurry image and create a new version that looks like a high-quality photo.

•In short, the project improves the quality of images by increasing their resolution and making them look more natural and visually pleasing.

•Our approach leverages the strengths of GANs to overcome the limitations of traditional super-resolution techniques by focusing not only on pixel accuracy but also on perceptual quality and visual realism.

Objective 

•The goal of this project is to create and train a Generative Adversarial Network (GAN) that can take low-resolution images—which are often blurry or pixelated—and turn them into high-resolution images that look much sharper and more detailed.

•This is useful in situations where the original image quality is poor, such as in old photos, security camera footage, or zoomed-in satellite images. Our aim is to make the images clearer and more realistic, so they look as close as possible to real high-quality photos.

•By training the GAN on lots of examples, the model learns how to add back the missing details in a smart way, instead of just guessing or stretching the image like traditional methods do.

Literature Survey

Block Diagram 

Model Description 

SRGAN (Super-Resolution Generative Adversarial Network) is a deep learning model designed to generate high-resolution (HR) images from low-resolution (LR) inputs.

Generator Network

Input: Low-resolution image

Output: Super-resolved image (same size as HR ground truth)

Architecture:

• Initial convolution layer with PReLU activation

• A sequence of Residual Blocks (each with:

• 2 convolutional layers

• BatchNorm

• PReLU

• Upsampling blocks using sub-pixel convolution (Pixel Shuffle)

• Final convolution to produce RGB output

Discriminator Network

Input: Real or generated HR image

Output: Probability of being real

Architecture:

• Series of convolutional layers with increasing feature maps

• LeakyReLU activations

• BatchNorm in intermediate layers

• Fully connected layers ending with a sigmoid for classification

Dataset Description 

• DIV2K stands for "DIVerse 2K resolution dataset"

• Created for image super-resolution tasks

• Introduced by the Computer Vision Lab, ETH Zurich

• Widely used in competitions like NTIRE SR Challenge

  •  Contains 800 high-resolution (HR) training images

• Image format: .png

• Resolution: ~2K (e.g., 2040 × 1080 or higher)

• Used as ground truth in super-resolution training

Results