Making photos looked scanned and photographed

[Neil284230581p7m]

Hello all

I am just writing as I have a batch of approximately 20,000 photos in png format. I would like to make it look like they've all been 1) scanned and then 2) been photographed.

Just writing to ask - does anyone know of any automated ways to do this?

Kind regards

Neil

[c.pfaffenbichler]

I am not sure what you mean.

• A good scan should not cause significant loss of quality. (Though naturally some loss of quality is unavoidable in the transfer.)

• What should have been photographed? A print, the screen, …? 

Couls you post examples and one or two of your unedited images? 

[Neil284230581p7m]

 The idea would be to take a photo like this and make it look like it had been scanned or photographed. I'd want to do this to 20,000 images

Something like this twitter feed: @ecgrhytyms (@ecgrhythms) / Twitter

As you can see the ECGs are all photographed in different ways with different amounts of light etc. The aim would be to make it look like someone has individually scanned all 20,000

Then the plan would be to make it look like someone has photographed all 20,000

[Chuck Uebele]

Yes, an example would be great. There are so many things that could be done. When I read this, I got the impression that you might want an old photo look of fading or cross processing, which  ould be done with curves, then for easier application, turned into LUTs.

[c.pfaffenbichler]

I wonder if the intended effect might go beyond colors/contrast and affect detail, too, (maybe even screens, lines, …) so an example would seem useful indeed. 

[jefbr]

Yeah same. I was thinking of actually scanning a couple of blank pages, each with varying wrinkles, storing them as texture and applying them in a randomly altered form per image.

[c.pfaffenbichler]

Good one, paper texture is another possible option. 

[Neil284230581p7m]

Thank you all for your help. Please see a clean version of one of the images.

The aim would be to make all 20,000 of them look like these: ECG interpretation in electrolyte emergencies | ECG Cases | EM Cases (emergencymedicinecases.com)

e.g. adding different amounts of noise, rotations, light views

What often happens in real life is these heart traces are printed on paper. Someone then takes a photograph or scans it to the electronic patient record. The aim would be to make it seem like this happened with all 20,000 of them.

[jefbr]

This is raising some red flags for me. 

Please hear me out on this:

If I understand you & the twitter feed correctly, I'm guessing that these 20k images will be used to train an AI model.

If that is correct*, I want to ask of you to help us think carefully about the kind of alterations you're looking for.

Changing a large batch of images "willy nilly" can introduce repeating patterns. (waves, noise patterns, reoccuring wrinkles of scanned paper, ... )

The model could treat these repetitions as significant if they regularly change the ECG line itself.

As an example carelessly applied alterations could perhaps introduce signs of fluttering/fibrilation.

I'm not a doctor, not even close, but I hope this example illustrates the kind of potential risk I'm talking about.

Am I overthinking it?

How do you estimate such a risk and why?

What will this be used for?

*It could just as well be that your supervisor is asking you to print & scan 20k images just to keep you busy.

In which case: let's automate it straight away and get you back to saving lives! 😉

[Neil284230581p7m]

Thank you for question. You are indeed correct. The plan would be to train AI algorithms using the dataset.

Your point about introducing repeat patterns is excellent as well. Thank you for thinking this through. You've certainly highlighted an important potential issue with this. Once we have created the ECGs, our plan would be to validate our dataset by comparing our created ECGs with the originals to ensure that they remain accurate from a diagnostic perspective.

Happy to discuss this in further detail.

[c.pfaffenbichler]

A superficial »dirtying up« of the image is possible, but as @jefbr mentioned it seems possible that this could introduce unintentional (and hard-to-notice) regularities that would not apply to the real-world-examples. 

Especially randomizing the result of simulated perspectival distortion and »wavy« paper might prove difficult.