Super-Resolution GAN for Image Upscaling

A PyTorch implementation of SRGAN (Super-Resolution Generative Adversarial Network) for upscaling low-resolution images from 64×64 to 125×125 pixels with enhanced perceptual quality.

Overview

This project implements a deep learning approach to single-image super-resolution using:

• Generator: 16-block residual network with channel attention mechanisms
• Discriminator: PatchGAN discriminator for realistic texture generation
• Two-stage training: MSE pre-training followed by adversarial fine-tuning
• Dataset: 111,000+ training images for robust learning

Key Features

• Residual blocks with batch normalization for stable training
• Channel attention modules for adaptive feature refinement
• Comprehensive metrics: MSE, PSNR, SSIM
• PyTorch 2.0 compatible with torch.compile() optimization
• Early stopping to prevent overfitting
• Checkpoint saving for best model preservation

Architecture

Generator

• Initial 9×9 convolution layer
• 16 residual blocks with PReLU activation
• Channel attention at blocks 8 and 16
• Pixel shuffle upsampling (2× scale)
• Bilinear interpolation for final resize to 125×125

Discriminator

• VGG-style architecture with strided convolutions
• LeakyReLU activation and batch normalization
• PatchGAN output for local realism assessment

Training Strategy

1. Phase 1: Generator pre-training with MSE loss for pixel-perfect baseline
2. Phase 2: GAN training with combined adversarial and perceptual loss
   • Generator learning rate: 1e-5
   • Discriminator learning rate: 1e-4
   • Early stopping with patience of 8 epochs

Requirements

torch>=2.0.0
torchvision
pillow
scikit-image
tqdm
numpy

Usage

# Load pretrained model
generator = Generator().to(device)
checkpoint = torch.load('best_generator_gan.pt')
generator.load_state_dict(checkpoint['generator'])

# Upscale image
with torch.no_grad():
    hr_image = generator(lr_image)