[rhousham]

Hi I've been given a very tricky project to do and was wondering if I could get some help.

Right the problem is that I have a website with 1000000+
Images and need a way to rate them in quality. Users vote and comment on the images so we think that there is a link between the users habits and the quality of the photo. It would be easy if users voted on the very good and only some voted on the bad. What we find tho is that some people click on their friends and some very good photos get overlooked.

I have just thought of implementing a neural net to solve the problem.

I just need some help to make it the best I can.
My idea is to make one neural net that attempts to rate how crediable a users vote is.

Using that I can then give the photo's a quality rating.

I just need some help about what to input and what to do.

It's just a mammoth task and it's been a while since I touched a neural net. Any help anyone can provide would be great.

Even if you think I have something would be great.
So I'm not barking up the wrong tree.
Richard

[one]

Richard,
things like credibility are subjective.....I mess around all the time but the computer deals in absolutes their is no way of knowing
maybe a rating like the effort to post is something but,
their are many things left out of your post...

If I remember I will try to understand and post in a couple of days..

Regards,
J.

[Art]

Rating your images with regard to quality needs some defining.

If subject matter is of no consequence then you are talking about pure
photo dynamics like, framing, centering, light, darkness, contrast, blur,
tonal quality, degree of color, tint, separation, etc., etc.

It matters not whether the subject is a dog, birdbath, leaf, raindrop,
etc., provided it meets your criteria of quality.

I think that would be an extremely tall order for a computer system
even one equipped with a great vision system and a huge database.

Users' habits are also guided by their individual likes and dislikes or
subject matter regardless of how subjective they believe they are.

What you're asking while perhaps, not impossible, is certainly
improbable.

Good luck no matter!

[one]

I can't say much, you took the words out of my mouth...

Q: what do you possibly want from/do with 1000000+ pictures?

[rhousham]

Hi lets try and explain this I have a site with loads of users loads of photos. Each member can vote on a photo. The thing is that you would expect the best images getting the top votes and the average images getting low votes. This does happen to some extent but what seems to happen is that some members get some friends and basically do vote groups. So because you have voted on mine I will vote on yours so there are people with average portfolios with high votes and there are some with good portfolios with medicore votes.

Hence the problem. My task is to somehow right the balance and itentify when these vote groups appear and attempt to give each photo a quality rating using the users habits as a factor.

I've given this some thought and it looks to be that the way to do it is to form maybe 2 neural nets(NN)
The first one taking the users average time on the site and average votes and several other factors and seeing how much of the vote grouping they do. So that will give us a value for each user

The second NN will take that value, information about the photo's owner and attempt to work out a quality rating of the photo.
What do you think now?

If the NN approach is the wrong way to go any other ideas.?
Ta
Richard

[Art]

Richard,

Perhaps you could configure it so that the users do not or would not know who each photo / picture belongs to or who submitted it. Basically, make the photo anonymous for all purposes. You could still keep an index (a large one at that with a million photos), of
the original submitters but not display them for others to see.

This would help eliminate the "buddy system" or voting for your friend's photo.

While the term Neural Net sounds nice a well structured database would do the job as easy. Don't over complicate things...Use the KISS principle whenever possible.

Regards!

[rhousham]

Thanks for the reply. I would love to make the photos owner annonomous but it's not been that way for several years so cannot really do that. Do you think tho that a NN could do the trick. Or have you another way.
Thanks
Richard

[one]

Richard,
>>"This does happen to some extent but what seems to happen is that some members get some friends and basically do vote groups. So because you have voted on mine I will vote on yours so there are people with average portfolios with high votes and there are some with good portfolios with medicore votes."<<<

Kill 'em All.

tough decisions, and I can't help, sorry.

Regards,
J.