AI Image Enhancement: Denoising Autoencoder (Proof of Concept)

A small proof of concept for the kind of work described in the job post: training a model to improve image quality, here specifically removing noise.

What's in here

• denoise_poc.py — the full thing, self-contained:
  • Generates its own training data (simple shapes with Gaussian noise added), so there's no dataset to source or download
  • A tiny convolutional autoencoder, under 10,000 parameters, built in PyTorch
  • A classical baseline (Gaussian blur) trained on nothing, just for comparison
  • Trains in a few seconds on CPU and plots the results
• results.png — clean vs noisy vs classical baseline vs trained model, side by side
• loss_curve.png — training loss over steps

Run it

pip install torch numpy opencv-python matplotlib
python denoise_poc.py

What it shows

On this synthetic data, the trained model brings the reconstruction error down to about a third of the classical baseline's (0.003 vs 0.010 MSE), and the output grid backs that up: the model keeps edges sharper where the blur just smooths everything out.

This is scaled down on purpose, small model, synthetic data, short training run, to stay lightweight and demonstrate the approach rather than solve a specific dataset. The same architecture and training loop extend directly to real images: swap in actual training data, size the model to the problem, and add augmentation or a stronger loss (perceptual loss, SSIM) depending on what "image quality" needs to mean for the actual use case.