Add Bad Photocopy Texture Augmentation

[sparkfish]

Add Bad Photocopy Texture Augmentation

[shaheryar1]

Yes, but the blobs should be in a diminishing/increasing pattern. It can't be totally randomized over the whole page

[proofconstruction]

This effect can occur when the page doesn't make adequate contact with the scanner bed, or when the glass of the scanner bed has defects or dirt blocking the light path

I think this can be achieved with existing noise and inkblobs augmentations, if they're restricted to sections of the image. Maybe a brightness texturize for good measure.

[kwcckw]

Yea, i'm trying to create different patches of noises using _create_blob and applyBlob:

But it should be better with blob size >1, right now those noises are in pixel level, each of them are having size==1, doesn't looks good yet.

[kwcckw]

I added variable size blobs, looks better now:

Yes, but the blobs should be in a diminishing/increasing pattern. It can't be totally randomized over the whole page

@shaheryar1 Yes, right now the example adding noises to 4 corners of image, so we would need to design different kind of noise pattern to make it more realistic.

Maybe a brightness texturize for good measure.

@proofconstruction I don't quite get this, could you elaborate further ?

[proofconstruction]

Maybe a brightness texturize for good measure.

@proofconstruction I don't quite get this, could you elaborate further ?

I just mean that the brightness of dots should also vary, in addition to their size. The bottom left corner of the original image has areas with a lighter form of the effect, which we could simulate by changing the brightness of the dots we apply in some area. This would give more variation to the effect than just the gaussian blobs, which look too "regular"

[kwcckw]

Maybe a brightness texturize for good measure.

@proofconstruction I don't quite get this, could you elaborate further ?

I just mean that the brightness of dots should also vary, in addition to their size. The bottom left corner of the original image has areas with a lighter form of the effect, which we could simulate by changing the brightness of the dots we apply in some area. This would give more variation to the effect than just the gaussian blobs, which look too "regular"

Okay, but shouldn't be photocopy machine output is always black & white? So the intensity variation might not be relevant here.

[kwcckw]

I added several patterns for this augmentation effect:

1. Scattered noises:
   

2. Random noises:
   

3. Corner noises:
   

A combination of them would looks like this:

I think the challenging part is to make them looks natural, so any comment or suggestion based on the current output above?

[kwcckw]

I added the colab notebook demo here:
https://colab.research.google.com/drive/1zcLdFEikxrQRQLFFEXiqnw9efevldrid?usp=sharing

[proofconstruction]

Merged PR #50 solving this.