Art : AI vs Human IV Library

 

Hello everyone! I hope you're doing well.

This is round 4 of my previous posts, as it is going to distance. I will start with a short summary of what was mentioned in my previous posts.

This series is a discussion on how AI itself can detect AI art from humans' art.

In the last post, we've seen how we can strip a painting into shapes, circles, lines, and boxes using a simple method called edge detection. Not one size fits all, but a good start for simple color schemes.

In this post, we will discuss what we will do with these shapes based on the hypothesis that we can turn all artworks into shapes. For more information, check out my last post.

So the idea is to use a library of images that we are sure were made by humans and contain all sorts of circles, lines, and squares. Then we would take our stripped artwork and use some of those good-old image manipulation techniques to compare them both. We won't get too deep into these techniques, but we can use convolutional neural networks (CNN) and train the model to spot features that are unique to those made by humans. We can use one-shot learning or transfer learning if the quality of our base images is not that great. We can limit them to those with good quality. At the risk of having a limited dataset, we may lose some accuracy (generally speaking), but the time consumption wouldn't be as heavy as with a massive dataset.

But that's not a problem in our case since the dataset of base images is made by yours truly, a great artist if I say so myself, and along the years I spent practicing this craft I made more drafts and scribbles than I made artworks and that left me with 4 sketchbooks with hundreds of pages filled with just circles, lines and squares.

Nonetheless, that leaves me with a rich library of drafts that can be used just for the purpose of this mission.

We will be focusing mainly on spotting the inconsistencies that we can see in hand-drawn shapes, such as a misaligned line, a wobbly circle, or a distorted square. On the other hand, we will be trying to spot consistent shapes, as they are the trademark of an AI model. Again, this idea doesn't work on all paintings, as I discovered that if an artist makes it his sole mission to make consistent shapes, he can pull it off as hard but possible.

This is the last part of theoretical speaking; hopefully next time we will put this model to the test and see the results we get.