Low-Light Image Enhancement

This repository contains all the necessary code and resources to enhance low-light images by reducing noise and converting them into high-light images. Our approach combines advanced neural network techniques with data preprocessing strategies to achieve superior image quality.

Overview

In this project, we aim to improve the Peak Signal-to-Noise Ratio (PSNR) of low-light images. We have refined our neural network architecture with batch normalization, which helps stabilize and accelerate the training process. Additionally, we utilized histogram-based preprocessing techniques to generate more effective input images, leading to substantial improvements in PSNR scores.

Table of Contents

1. Dataset
2. Exploratory Data Analysis and Preprocessing
3. Model and Architecture
   • Preprocessing Layer
   • Special Convolutional Module
   • Output Layer
   • Main Model
4. Losses and Training
5. Usage

Dataset

The dataset consists of 485 sets of paired images taken in both low-light and high-light conditions. Each set includes:

• A low-light image
• A corresponding high-light image

Exploratory Data Analysis and Preprocessing

We begin by analyzing the percentiles of low-light and high-light images to identify significant correlations (≥ 0.6). Our preprocessing involves quantile regression models for histogram mapping, using an XGBoost regressor trained on the histogram data of low-light images.

Model and Architecture

Preprocessing Layer

The Preprocessing Layer applies initial convolutional operations to the input data, using a single convolutional layer with a 3x3 kernel size, followed by a ReLU activation function.

Special Convolutional Module

This module processes data through two stages:

1. Two pathways in the first stage:
   • A 1x1 convolutional layer
   • Two 3x3 convolutional layers
2. Combined pathways in the second stage:
   • Two 3x3 convolutional layers
   • A shortcut connection (residual learning)

Output Layer

The Output Layer generates the final feature representation through a single convolutional layer with a 3x3 kernel size and a sigmoid activation function.

Main Model

The Main Model includes:

• An initial layer
• A series of special convolutional modules with skip connections
• A final layer with sigmoid activation

Losses and Training

We employ a combined loss function minimizing MSE and MAE scores. Training is optimized with:

• AdamW optimizer
• Learning rate scheduler (ReduceLROnPlateau)
• Mixed precision training with torch.cuda.amp.GradScaler
• DataLoader for batching
• Early stopping based on validation loss

After training for 5 epochs, we achieved a PSNR score of 24.57.

Usage

To use this repository, follow these steps:

1. Clone the repository:

   git clone https://github.com/yourusername/low-light-image-enhancement.git

2. Install the required dependencies:

   pip install -r requirements.txt

3. Run the main script:

   python main.py