Low-Light Image Enhancement with Deep Learning
Overview
This project focuses on developing a deep learning model that enhances images taken in low-light conditions while preserving important details and avoiding over-exposure artifacts.
The Problem
Images captured in low-light environments suffer from:
- Low visibility and poor contrast
- High noise levels
- Loss of color information
- Difficulty in identifying important details
Our Approach
We developed a CNN-based architecture that:
- Analyzes the illumination map of the input image
- Separates noise from actual image content
- Enhances brightness while preserving natural colors
- Restores fine details lost in dark regions
Key Technologies
- PyTorch: Deep learning framework
- CNN Architecture: Custom encoder-decoder network
- LOL Dataset: Low-light image pairs for training
- OpenCV: Image processing and evaluation
Results
Our model achieved:
- PSNR: 24.5 dB on test set
- Processing Time: 50ms per image on GPU
- Visual Quality: Natural-looking enhancements without artifacts
Published at IJCNN 2025 as “RSEND: Retinex-based Squeeze-and-Excitation Network with Dark channel prior”
Before/After Examples
The model successfully enhanced cathedral interiors, street scenes, and indoor photography while maintaining natural color balance and avoiding the “washed out” appearance common in traditional enhancement methods.
Future Work
- Real-time video enhancement
- Mobile device optimization
- Adaptive enhancement based on scene content