Create "Damage Paper" Augmentation

[proofconstruction]

Motivation
It's not uncommon to scan documents that have undergone some physical deformation, like tearing, folding, or crinkling. The resulting changes in the surface of the paper generally become more apparent after digitization, causing difficulties for humans and machines reading the text.

See, for example, the image at the top of this post

It would be useful to be able to generate images of "damaged" documents, for training models in settings like healthcare (medical records), law (contracts), finance (invoicing), and so on.

There are several forms of damage that could be applied. Off the top of my head:

1. crumpling
2. tearing
3. burning
4. dissolving

... to name a few

Implementation
There are several paths forward for something like this. The most naive way would be to take some images of damaged paper and use these as the base image for existing pipelines. One much more sophisticated approach would be generating a 3D model (perhaps using Blender API?) and applying an image of a document as a texture.

[kwcckw]

One much more sophisticated approach would be generating a 3D model (perhaps using Blender API?)

Interesting, or we can just use some damaged paper reference in paper phase?

[proofconstruction]

Interesting, or we can just use some damaged paper reference in paper phase?

Can you explain more what you mean?

[kwcckw]

The most naive way would be to take some images of damaged paper and use these as the base image for existing pipelines

Actually i think this should be good enough, but how a 3D model can further improve the results? Do you have any related link on that?

[proofconstruction]

I don’t have a link, and I’m not sure what to look for.

What I mean is:

1. We use a pipeline to create an image, just like we can now
2. After the image is generated, we pass it through the CrumplePaperAugmentation to simulate rolling the paper into a ball, tearing it, etc.
3. The result is very different than using the existing image pipeline to apply text to a picture of damaged paper, because the text does not deform in the same way the paper is.

[jboarman]

Before tacking the more complicated concepts in this issue, we might want to decompose the concepts into simpler techniques (like in #21) that we can evolve into more the complicated techniques that we'll need for fuller implementation needed for this issue.

[jboarman]

The mechanical deformations proposed here are good ones that should be independently proposed on a per-transform basis along with examples of the expected output. For now, I think we should allow issue #21 (single paper fold) to reach completion so that we can pivot accordingly at that point based on how that issue matures.