Can machine learning tell you if your pictures are awesome?

Everypixel have recently launched a web app EveryPixel that uses machine learning to examine your images and tell you if they are "awesome" and how awesome they are!

You will know from my earlier posts on this topic that I find the whole work going on in artificial intelligence and art in general, photography in particular to be a very fascinating field of development right now. This kind of picture assessment, beyond just identifying objects and people within a scene is another step along that path.

What the developers at Everypixel have done is taken an absolutely huge collection of images (just shy of a million) and asked designers, editors, photographers and others to rate the images. They have also hooked this up with their stock photography business to compare purchases and value too.

It is a really interesting idea. Can a soulless machine actually learn how to discrimate artistically and "see" the beauty of images the same way a human does?

They have a good case for doing so, these days stock photographer is awash with images, for stock image buyers weeding the trash from the excellent is time consuming and painful. So their aim is to show stock images not just by keyword but by quality too.

There are a range of crowd source systems that sort of do this already - I have my own at which lets you upload your photos and then vote

Sites like Purpleport with it's FPI's and daily competitions and 500px with it's "popularity" rating are other ways of letting humans do the work.

We've also seen activity from companies like Adone who are trying similar machine / AI approaches to image assessment.

So, does this really work? Surely every image is unique and although we often talk about the "rules" of photography, many images are still powerful or beautiful without following any of those so-called rules?

I think that is half true. The point about contemporary AI systems is that they are no longer coded by programmers with a set of rules - instead they are told this is a good image, this is a bad - you decide what in the image makes it worth looking at.

It is possible for such systems to develop their own ways of seeing - they may look at flows with the image, balance of colours, areas of contrast - or maybe something much more subtle that humans haven't really identified.

So, here's the acid test - how well does it perform? I banged a selection of images into the system to see what it came back with.

You can see it gives a percentage "awesome" then categorises the contents of the image.

I think 100% is reasonable for one of the most iconic images of the 20th century!

So, I'm going to show you 10 other images of mixed quality, you decide how they should be rated then I'll give you the results.

OK, got your views in mind - let's see what the EveryPixel machine came up with...

96% - OK. It's a striking picture, strong colours and lines, not sure it's worth that much though.

0% - this is just one instance of 5 that it just failed to score at all. I have not bothered putting the other zeros in. I think there are some images where it just fails to analyse rather than actually thinking they are 0 score pictures.

27% - Bill Brandt's image - I sort of agree. It's not one of his best structured images, but it is very striking. I think I give the benefit of the doubt to the machine here.

100% - Can't argue with that - I have to say I am very impressed with its rating here - this is a very busy picture and knowing your history seems to be important too. So, whatever criteria the machine is using it has seen something in the image itself without that historical knowledge.

41% - Nick Knight - a contemporary iconic image - I think it has scored this much lower than it should be. Not sure why. You could make arguments about the way the body fits (and slightly cuts into) the frame may be factors that let it down.

28% - Mine when I first started. Not an unreasonable score - but I think humans would rate it higher.

97% - Mine. I think this is a striking picture, the lines and the colour - but I would rate it down in the 50's.

18% - Mine. This is a deliberately "bad" image, what I mean is I've desaturated it, left it crudely posed and junk in the background as part of the artistic intent. So, I agree but would give myself more points because "art".

20% - Mine. Not sure - I think this is maybe a 40's / 50's image.

99% - Mine. Very interesting. This isn't an image of mine that I personally rate very highly. It's perfectly fine - just doesn't seem to me to offer anything special. However it is one of my most popular images - so maybe the machine knows better.

How did your results compare? Did you guess with me or the machine? Was it close enough?

We have to be fair here, anyone who has seen a photo judge in action will know that any human judge will also rate images very widely and often subjectively under- or over-score images.

Maybe it is no worse than a typical judge.

If you want to read more about Google and Microsoft and their work in AI vision this article - AI Vision Detecting Nudes may interest you.