XLSR - Extremely Lightweight Quantization Robust Real-Time Single-Image Super Resolution for Mobile Devices

Welcome to the Pytorch-Lightning implementation of the research paper "Extremely Lightweight Quantization Robust Real-Time Single-Image Super Resolution for Mobile Devices (XLSR)". XLSR is a cutting-edge solution engineered to deliver exceptional image super-resolution while remaining lightweight, quantization-robust, and optimized for real-time execution on mobile devices.

---

Features

• Lightweight Model: Designed for minimal memory footprint and computational efficiency.
• Quantization Robustness: Ensures high performance even after quantization, enabling seamless deployment on mobile hardware.
• Real-Time Performance: Provides low-latency performance, ideal for mobile and embedded systems.

---

Table of Contents

1. Installation
2. Usage
3. Model Architecture
4. Dataset
5. Training
6. Results
7. References

---

Installation

Requirements

To get started, ensure the following dependencies are installed:

• Python 3.10 or later
• PyTorch 2.3 or later
• PyTorch Lightning (latest version)
• Pillow
• NumPy

---

Usage

Quick Start Example:

see test.ipynb

---

Model Architecture

XLSR leverages an efficient design composed of:

• Efficient Residual Blocks: Enhances feature extraction with minimal overhead.
• Pixel Shuffle Layers: Upscales images efficiently with reduced computational complexity.

---

Dataset

XLSR supports the following datasets:

• DIV2K - A high-quality image super-resolution dataset.
• Place the dataset into the dataset directory, with training images stored in the DIV2K_train_HR directory and validation images in the DIV2K_valid_HR directory.

---

Training

To train the XLSR model, execute:

python main.py

---

Results

Benchmark Results

Model	Parameters (K)	PSNR (dB)
XLSR	20	29.58

---

References

• Original Paper: arXiv:2105.10288
• Journal reference: IEEE Computer Vision Pattern Recognition Workshops (Mobile AI 2021 Workshop)