Low-Light Image Denoising Project

Overview

This project focuses on enhancing low-light images by reducing noise and converting them into high-light images. The main metric used for evaluating the improvement is the Peak Signal-to-Noise Ratio (PSNR).

Dataset

• Total Images: 485 sets of paired images
• Structure: Each set includes one low-light image and one high-light image

Exploratory Data Analysis (EDA)

• Plotted percentiles of low-light and high-light images
• Significant correlations (≥ 0.6) observed between percentiles
• Quantile regression model and XGBoost regressor used for histogram mapping

Figure 1: Histogram Comparison

Key Techniques

1. Preprocessing: Initial convolutional operations to extract features
2. Special Convolutional Module: Inception-like structure with added pathways and residual learning
3. Output Layer: Final convolutional operations followed by sigmoid activation

Figure 2: Model Architecture

Training

• Loss Function: Combination of Mean Squared Error (MSE) and Mean Absolute Error (MAE)
• Optimizer: AdamW with learning rate scheduler
• Early Stopping: Stops training if validation loss does not improve for 5 epochs
• Mixed Precision Training: Utilizes torch.cuda.amp.GradScaler for performance

Results

• Achieved a PSNR score of 26.25 after 5 epochs

Figure 3: Image Enhancement Results

Installation

To install the required packages, run:

pip install -r requirements.txt