Examples of different approaches of doing noise removal with neural networks.

The approaches are good for situations where you don't have any clean/denoised imagery to work with. Examples of that can be weird sensors, old photos, files with artifacts from compression and/or hardware. I have not gotten around to actually testing this stuff with a decent dataset, so I am not really sure how the results are. Thus far the 1555 lines of code is mainly made with mixtral, gpt-4o, codestral and some meat input. But in the future this approach is probably gonna change to be a mix of different models with a main model using langchain as a broker that will utilize codestral and gpt-4o and do lookups of various sources on the internet plus it's own repo as the main context to speed up development. Meat doing code is a bad approach for sure - so I'll move away from that as the toolchain progresses.

Make sure you have cuda 12.3 and libcudnn-frontend-dev/mantic,mantic 0.8+ds-1 all + nvidia-cudnn/mantic 8.9.2.26~cuda12+1 amd64 ... maybe other things as well.. Not really sure..

git clone yadayada....; cd yadayada.... # then make the venv:

python3 -m venv .

source ./bin/activate

./bin/pip install -r requirements.txt

./bin/pip install opencv-python .. since I have not added it to the thing

By this time you should have a good setup for playing with this stuff without messing up your base py setup.

Then to quickly test your setup is working properly:

to test pytorch:

python3 test-pytorch-setup.py

to test tensorflow:

python3 test-tensorflow-setup.py

Then we have a couple of ways to denoise:

Autoencoders (tensorflow): Should work fine'ish.

Noise2Noise (pytorch): A little better - but it's a classic CNN.

CycleGAN (tensorflow): The most promising of the three.

BlindSpot (pytorch): Another approach using CNN. Very dramatic results.

cGAN (pytorch): Conditional GAN. No idea if it works.

Visualize: A tool for comparing all the four images.

Tracker: A very basic resnet tracker for doing bounding boxes on interesting stuff in the images.

AutoRegressor: Train on training data and generate synthetic data using autoregressor.

Note that this is just example code and needs a lot of adaptation for them to work on your datasets.

All of them expect input in data/training and will output in their own directory under data.