Comparative Analysis of Image Super-Resolution Models

A comprehensive analysis of three deep learning models for image super-resolution (SR): Enhanced Deep Super-Resolution (EDSR), Very Deep Super-Resolution (VDSR), and Deep Recursive Convolutional Network (DRCN).

Overview

This project compares the performance of three popular super-resolution architectures with different design philosophies:

• EDSR: A deep residual network architecture with 32 convolutional layers
• VDSR: A deep network with 20 convolutional layers focused on residual learning
• DRCN: A recursive architecture that leverages weight sharing across layers

Dataset

• Training: DIV2K dataset (800 training, 100 validation images in 2K resolution)
• Testing: Set5 benchmark dataset

Model Architectures

EDSR

• Deep residual network with 32 convolutional layers
• Uses MeanShift layer for input normalization
• Includes multiple Residual Blocks with ReLU activations
• Employs pixel shuffling for upsampling

VDSR

• 20 convolutional layers with 64 channels each
• Global skip connection for residual learning
• ReLU activation after each convolutional layer (except final output)
• Trained using L2 loss

DRCN

• Embedding network with two convolutional layers
• Recursive inference network that applies the same convolutional layers multiple times
• Reconstruction network that combines recursive outputs
• Custom loss function combining MSE, color loss, recursive loss, and L2 regularization

Training Setup

Model	Image Size (HR/LR)	Loss Function	Optimizer	Learning Rate	Epochs
EDSR	512x512 / 128x128	L1 Loss	Adam	0.0001	20
VDSR	128x128 / 64x64	L2 Loss	SGD	0.001	50
DRCN	512x512 / 128x128	Custom	Adam	0.0006	70

Results

Image Quality Metrics

Model	PSNR	SSIM	Perceptual Index (PI)
EDSR	30.14	0.8666	-5.5048
VDSR	29.82	0.9574	-5.3873
DRCN	31.86	0.8876	-10.3662

Computational Efficiency

Model	Parameters	Inference Time (ms)	Memory Usage (MB)
EDSR	43,089,923	4126.71	512.48
VDSR	668,227	136.63	6.02
DRCN	1,888,148	101.54	0.21

Key Findings

• DRCN demonstrates the highest PSNR, indicating strong reconstruction capabilities, with excellent efficiency for real-time applications
• VDSR achieves the highest SSIM and good PI scores, balancing structural and perceptual quality with reasonable computational demands
• EDSR provides balanced reconstruction quality but requires more computational resources

Future Work

• Explore hybrid approaches combining strengths of multiple architectures
• Improve training techniques for faster convergence
• Adapt models to diverse datasets for broader applications

Installation and Usage

# Clone the repository
git clone https://github.com/Lynn00Chen/image-resolution.git

cd image-super-resolution-comparison