noise2normal

Denoising Images using Central Limit Theorem

A tiny, practical demo of the Central Limit Theorem applied to images. We generate multiple noisy versions of the same image and average them to recover a cleaner result.

It shows that regardless of the noise distribution you start with (uniform, skewed, or already Gaussian), the sample mean of independent noise terms trends toward a Gaussian around the true mean, with variance shrinking like 1/n. In code here we use Gaussian noise by default, but the averaging logic and takeaway generalize.

Build

g++ -std=c++14 main.cpp -o noise2normal $(pkg-config --cflags --libs opencv4)

Run

./noise2normal -i <image_path> [--sample-size N] [--show-random-noisy-image]

Args:

• -i, --image <path> input image path (required)
• --sample-size N number of noisy samples to generate and average, default 5
• --show-random-noisy-image also show one random noisy sample window

Example

./noise2normal -i sample.png --sample-size 1000 --show-random-noisy-image

Noisy sample	After CLT (N=1000)