Facial recognition algorithms are often biased because the training data doesn't contain enough diverse faces.

What it does

The solution uses StyleGAN2 to create additional samples to train models on.The tool target group are data scientists, machine learning engineers and developers. Having more diverse images would prevent future face recognition/detection algorithms from being biased towards certain ethnicities.

How I built it

I used NVIDIA's research paper on StyleGAN2 algorithm and used it to generate artificial images. My version let's anyone with a browser use the solution and download the generated images in bulk

Challenges I ran into

I don't have a GPU on my personal computer so I had to use Google Colab

Accomplishments that I'm proud of

I managed to make a fully functioning minimum viable product

What I learned

I learned about the theory and applications of general adversarial neural networks. I learned hoiw to use the Tensorflow framework

What's next for GAN Face images sample generator

I'd like to build a web app for the solution and deploy it on Heroku or pythoneverywhere so anyone could easily access it from the browser

Built With

  • colab
  • numpy
  • pillow
  • style2gan
  • tensorflow
Share this project: