[Research] Implement Deep Learning Spatial Denoiser (DnCNN / GAN) to replace classical NLM

[TechieSamosa]

🛑 Research Track Issue

Please read RESEARCH.md before proceeding. Do not submit a Pull Request for this issue without first commenting a brief proposal and receiving approval from the core team.

The Problem:
Currently, after our core Zero-DCE model enhances the faint lunar light, we use a classical Non-Local Means (NLM) algorithm to clean the amplified sensor noise. While effective, NLM is slow and mathematically rigid.

The Research Goal:
We are looking for an ML Engineer to design and train a dedicated Deep Learning denoiser (e.g., DnCNN, FFDNet, or a lightweight GAN) specifically tuned for mixed Poisson-Gaussian (MPG) sensor noise.

What you would do:

1. Generate synthetic noisy/clean pairs using our OHRC noise profile.
2. Train a standalone PyTorch denoising model.
3. Integrate this model into src/inference/pipeline.py as a replacement for the cv2.fastNlMeansDenoising step.

To claim this: Comment below with a 3-4 sentence proposal of which architecture you would use and why it is suited for heavy sensor noise.

[TensorNaut]

I’d like to take this research task by implementing a DnCNN-based denoiser trained specifically on synthetic noisy/clean image pairs generated from the OHRC mixed Poisson-Gaussian noise profile. DnCNN is a strong fit because its residual-learning formulation is highly effective at separating signal from heavy sensor noise while remaining lightweight enough for efficient inference within the existing enhancement pipeline. Compared to classical NLM, it can provide significantly faster runtime and better adaptability to the non-linear characteristics of amplified lunar imaging noise. After training and evaluation, I will integrate the model into src/inference/pipeline.py as a drop-in replacement for the current cv2.fastNlMeansDenoising stage.

[TechieSamosa]

Assigned