Color-Sensation

Left: Black and White. Right: Colorized
Left: Black and White. Middle: Original. Right: Colorized
Top Left: Black and White. Top Right: Colorized. Bottom: Original

Inspiration

There was a Vox video on colorization, and we were interested in seeing how well a neural network could generalize and do this. We wanted to see how difficult it was for these artists to remake these images and how much time they put in. We also wanted to see if we could potentially help out with the process by creating a program that could create a baseline for color restoration and make a process that normally takes weeks into one that only takes a few hours.

What it does

We put a series of training images in the /Training/ folder, and then the neural network runs. It then goes to the /Testing/ folder and colorizes the photos in there, giving us results. We were planning on including a web app that takes in images and runs the neural networks, but we were unable to complete it. It uses primarily landscapes, forests, and grass due to limitations in its ability to generalize to vibrant colors.

How we built it

We used Keras and Tensorflow to build the primary engine and neural networks. We looked at a few articles to learn about basic syntax and some optimal structures for neural networks.

Challenges we ran into

Keras originally gave us a lot of issues since we were uncomfortable with the software. We had to look at various resources to fix issues. Additionally, we had to deal with issues like We also had many troubles with the web app, including uploading images, displaying them, and running the neural network. Additionally, the simple structure that we used for the neural network did not do a great job generalizing to different types of images. When there were different, vibrant hues, it would blend colors together, which could be fixed with more advanced structures.

Accomplishments that we're proud of

We managed to create a few good images. We were also able to generalize pretty well to forests and grass scenery. Overall, as none of us knew Keras or Tensorflow going into the event, we were pretty happy to have at least some results that we are proud of.

What we learned

How to use Keras/Tensorflow at a basic level and the limitations of basic neural networks

What's next for Color-Sensation

We want to hopefully finish the web app and get a better interface for our application. We also want to be able to perhaps add more complicated structures to let the neural network generalize better to unknown data.