SRGAN.

Overview

SRGAN is a computer vision project exploring image super-resolution with a generative adversarial network. It takes a low-resolution image and learns to reconstruct it at a higher resolution.

The project examines the relationship between reconstruction fidelity and visual quality. A sharper-looking image may not reproduce the original details exactly, so evaluation needs both quantitative metrics and visual inspection.

Problem and objectives

Simple interpolation often produces smooth or blurred details when enlarging an image. Super-resolution models learn visual patterns from image pairs to produce a richer reconstruction.

The project has three objectives:

• Implement a generator and discriminator for single-image super-resolution.
• Explore residual learning and upsampling for image reconstruction.
• Prepare evaluation that considers structural similarity and perceptual quality.

Model architecture

The generator starts with a 9 x 9 convolution and PReLU activation, followed by residual blocks and a long skip connection. This structure preserves input information while learning features for reconstruction.

PixelShuffle increases spatial resolution, and the output layer forms the reconstructed image. The base configuration uses eight residual blocks and 64 feature channels.

The discriminator uses convolutional blocks with increasing feature capacity, downsampling, adaptive pooling, and a classification head. During adversarial training, it provides a learning signal about the visual characteristics of generated images.

Experiment preparation

The experiments use DIV2K for super-resolution research. Preparation includes low-resolution inputs and high-resolution targets, scale configuration, and a separate evaluation split.

The generator supports scale factors of 1, 2, and 4, with a default of 2. The 4x configuration uses two upsampling stages. Scale factors and checkpoints must match for input-output comparisons to be meaningful.

The training approach combines adversarial learning with VGG-based perceptual assessment.

Quality evaluation

Evaluation is designed around:

• PSNR: measures pixel-level reconstruction error.
• SSIM: assesses structural similarity between reconstruction and target.
• Visual comparison: examines texture, sharpness, and artifacts that a single metric may not adequately describe.

Reproducible comparisons require consistent scale factors, degradation procedures, splits, checkpoints, and evaluation settings. This case study does not present benchmark scores as a completed final result.

Implementation outcomes and limitations

The implementation covers a residual generator, an adversarial discriminator, multiple scale factors, and a PyTorch experiment workflow.

Generated details are reconstructions based on learned patterns. Applications requiring exact visual information need additional review, because plausible details can differ from the original scene.

Next steps

The next priorities are measured comparisons against interpolation baselines, connecting each visual example to its input and checkpoint, and examining the balance between distortion metrics and perceptual quality.

Experiment documentation should retain training and evaluation settings so that model changes can be compared consistently.

Code and documentation

• Super-resolution project code
• Generator and discriminator implementation