Background
Image denoising can be viewed as a simple inverse problem: recover a clean image from an observed image corrupted by noise. This project starts with a deliberately small scope so the mathematical assumptions are easy to inspect.
Method
The MVP compares a baseline smoothing method with a regularized reconstruction objective. The main idea is to balance fidelity to the noisy observation with a penalty that discourages unstable or overly rough reconstructions.
Results
The early experiments show the expected tradeoff: stronger regularization removes more noise but can blur edges and fine details. Visual comparisons and simple metrics help make that tradeoff concrete.
Technical Stack
- Python for experiment scripts
- NumPy for array operations
- Matplotlib for visual comparison plots
- scikit-image for sample images and basic quality metrics
Next Steps
- Add total variation regularization
- Compare Gaussian, salt-and-pepper, and Poisson noise
- Write a short note explaining the optimization objective
- Build a small interactive demo for parameter tuning