MIRNet for Low-Light Image Enhancement

This repository implements and experiments with MIRNet, a deep learning architecture for enhancing low-light images. The project is inspired by the work presented in Learning Enriched Features for Real Image Restoration and Enhancement and was developed as part of the EE610: Image Processing course under the guidance of Prof. Amit Sethi at IIT Bombay.

---

Table of Contents

1. Introduction
2. MIRNet Architecture
3. Experiments Conducted
4. Results
5. Dataset
6. Installation
7. Usage
8. References

---

Introduction

Low-light conditions degrade image quality, resulting in issues like noise, low contrast, and color distortion. Enhancing such images is a challenging problem in computer vision. This project uses MIRNet, a multi-scale convolutional neural network that:

• Captures spatial accuracy.
• Reduces noise and preserves details.
• Improves perceptual quality using hybrid loss functions.

The work focuses on restoring high-quality content from low-light images, particularly using the LoL Dataset.

---

MIRNet Architecture

MIRNet introduces the following key components:

1. Multi-Scale Residual Block (MRB): Captures multi-scale contextual information while maintaining high-resolution details.
2. Dual Attention Unit (DAU): Combines channel attention and spatial attention for better feature refinement.

The network is trained using a hybrid loss function combining:

• Charbonnier loss: Robust loss for image restoration.

Optimizer: Adam
Learning Rate Scheduler: ReduceLROnPlateau

---

Experiments Conducted

1. Baseline Implementation

• Implemented MIRNet with original architecture and parameters.
• Loss Function: Charbonnier Loss.

2. Modified Architecture

• Experimented with Dual Attention Units (DAU) for refining features.
• Improved SSIM and PSNR metrics.

3. Alternative Loss Functions

• Added SSIM Loss alongside the Charbonnier loss to emphasize perceptual quality.\

4. Added regularization

• Experimented with the addition of L1 and L2 regularization

Each experiment can be found in:

• mirnet.ipynb: Baseline implementation.
• mirnet_modified.ipynb: Modified architecture.
• mirnet_modified_SSIM_loss.ipynb: Alternative loss functions.
• mirnet_modified_regularisation.ipynb: Added regularization.

---

Sample Outputs

Input	Enhanced (Baseline)	Enhanced (Modified)

---

Dataset

The LoL Dataset was used for training and evaluation. It contains paired low-light and well-exposed images:

• Training: 485 image pairs.
• Testing: 15 image pairs.

Dataset link: LoL Dataset

---

Installation

1. Clone the repository:

   git clone https://github.com/TheShiningVampire/MIRNET_for_low_light_image_improvement.git
   cd MIRNET_for_low_light_image_improvement

2. Set up the environment:

   python -m venv venv
   source venv/bin/activate
   pip install -r requirements.txt

---

Usage

To run the experiments, execute the following Jupyter notebooks:

1. Baseline Implementation:

   jupyter notebook mirnet.ipynb

2. Modified Architecture:

   jupyter notebook mirnet_modified.ipynb

3. SSIM Loss Implementation:

   jupyter notebook mirnet_modified_SSIM_loss.ipynb

---

References

1. Learning Enriched Features for Real Image Restoration and Enhancement
2. Keras MIRNet Example
3. LoL Dataset