Artificial intelligence makes blurry faces look more than 60 times sharper

[AnandGupta]

Read about it today https://techxplore.com/news/2020-06-artificial-intelligence-blurry-sharper.html

Now, since the code is available at https://github.com/adamian98/pulse, I think it is Python, can we make something in B4X ?

Regards,

Anand

 

[CaptKronos]

I'm not sure what you plan to use this for but note what they say:

The system cannot be used to identify people, the researchers say: It won't turn an out-of-focus, unrecognizable photo from a security camera into a crystal clear image of a real person. Rather, it is capable of generating new faces that don't exist, but look plausibly real.

If you want some fun then play away but I think for serious applications you will find this approach wanting.

 

[udg]

It won't turn an out-of-focus, unrecognizable photo from a security camera into a crystal clear image of a real person

It could be used the same way as an identikit drawing based on witness description. You can't have the exact face of the suspect but something near.
Training AI using a DB of known robbers will probably lead to good results.

Another use could be to try to identify some object in the scene that's not well defined. Yes, AI could be totally wrong, but afterall there's HI (humain intelligence) that should evaluate and process info from the former..hopefully..

 

[Sandman]

there's HI (humain intelligence) that should evaluate and process info from the former..hopefully..

Hopefully yes. But likely? Not so much.

This is about a post from a Googler, posted on Google+ (now dead) about why Google doesn't have customer service:
https://www.seroundtable.com/google-support-staff-limits-13916.html

Relevant quote:

If you have a billion users, and a mere 0.1% of them have an issue that requires support on a given day (an average of one support issue per person every three years), and each issue takes 10 minutes on average for a human to personally resolve, then you'd spend 19 person-years handling support issues every day.
If each support person works an eight-hour shift each day then you'd need 20,833 support people on permanent staff just to keep up.

That, folks, is internet scale.

So no, I'm not hopeful. Just look at how things are going with the automatic evaluation of apps on Google Play. It's not very simple to reach a human, and when you do they don't seem to have the authority to use any HI - they are just used as a warm meatbag following a very narrow script.

 

[AnandGupta]

I'm not sure what you plan to use this

Thanks for reply.

Actually I am interested only to make all those photos taken with small pixel size, old phones, which look pixelated and worse on bigger screen, to look good.
Many of us have old small photos which have our dream and wishes in it. I want to make them to look better on, say 40 inch screen.
For example, when we share photos on Whatsapp it makes small size of large photo, to the receiver, which gets pixelated on bigger screen.

Regards,

Anand

 

[Sandman]

Actually I am interested only to make all those photos taken with small pixel size, old phones, which look pixelated and worse on bigger screen, to look good.

I have to ask: You DO realize that if you use this solution, the end result is a good-looking image of a face that doesn't match the actual person?

I mean, just take a look at their own photos, from the page you linked:

The top row are their actual faces. Middle row is the pixelated version. And the bottom row is their re-created faces. Note that the top row and the bottom row does not match very well.

As far as I can tell their solution has no actual practical value beyond slight amusement. Perhaps it's a lack of imagination on my part, but I just don't see it.

 

[AnandGupta]

As far as I can tell their solution has no actual practical value beyond slight amusement. Perhaps it's a lack of imagination on my part, but I just don't see it.

Ahh.. we are taking about making an app and the result should be sufficient to make the owner happy to see his loved ones more clearly than the pixelated version. Human can recognize friends and family members even if they grow beards

Regards,

Anand

 

[CaptKronos]

Since it sounds like your requirement is a one-off activity I would suggest you just get it running in a regular desktop Python environment rather than going through (what I would expect to be) a difficult transition to B4X.

 

[AnandGupta]

Yes, I understand.
Just thought if I can make it an app then may be it will be helpful to users too.

Maybe next time.

Regards,

Anand

 

[CaptKronos]

Let us know how you get on!

 

[JackKirk]

You might like to also have a look at https://letsenhance.io/ which I am attempting to explore.

I have a use for something like this - to enhance reasonable images into more reasonable images...

 

[rabbitBUSH]

You DO realize that if you use this solution, the end result is a good-looking image of a face

At my age that's a GOOD thing.

Is there something they can use to dig out information from the image file which is there but not used because the resolution the display software uses ignores it? Because the suggestion is being pointed out here that this software 'manufactures' or fills in gaps by ALGORYTHM techniques hence its not a reality match at the end.

 

[rabbitBUSH]

to enhance reasonable images

Yes there is a use in many places. I have often used the phone camera to 'see' a label in an awkward place but it can still be a bit unclear. I've used it to see onto the depths of an engine bay at times and so on.

Cameras usually do a good job, but on dim light an enhancer would help.

But, that might be a different brand of software which I have found elsewhere anyway. Years ago I used an early digital camera {like one pixel res. days} to photo friends in a dark sheep barn and got a dark room effect. Ran some gamma corrections etc and was amazed to get people, dog and sheep appear. So the data was there in the image file.

Point being that maybe this is just a new version of the same thing GIMP etc can do. In actuality.

 

[ivan.tellez]

Actually I am interested only to make all those photos taken with small pixel size, old phones, which look pixelated and worse on bigger screen, to look good.
Many of us have old small photos which have our dream and wishes in it. I want to make them to look better on, say 40 inch screen.
For example, when we share photos on Whatsapp it makes small size of large photo, to the receiver, which gets pixelated on bigger screen.

You should first understand what this proyect is doing. It has NOTHING to do with upscaling or improving a pixelated image. It just generates random photorealistic but fake faces, downscale them and if this downscaled looks like the input, it gives you that image as a result.