This repository contains the implementation of a low-light image enhancement model that was used in the NTIRE-2024 Low Light Image Enhancement competition.
This model is based on the ImageLab architecture, specifically tailored for low-light image enhancement. This architecture integrates several techniques to improve the image quality:
- Spatial Information Enhancement: Improves the detail and structure of low-light images.
- Intricate Feature Capture: Effectively captures detailed features in low-light conditions.
- Multi-Scale Feature Refinement: Refines features at multiple scales for better image quality.
- Effective Noise Suppression: Reduces noise while preserving important image details.
The enhanced images produced by this model achieved an average Peak Signal-to-Noise Ratio (PSNR) value of 24 dB on the training set.
- Training Platform: Kaggle Notebook
- GPU Used: NVIDIA Tesla P100
The model architecture and training methodology are based on the research presented in the following paper: