Google Images: Visually Similiar

Started by grnidone, January 23, 2012, 08:29:41 PM

Previous topic - Next topic
Print
Go Down Pages1
User actions

grnidone

January 23, 2012, 08:29:41 PM Last Edit: January 23, 2012, 08:31:21 PM by grnidone
How does Google determine if an image is "similiar" to any one particular image?

I wanted to see if there were any more similiar to the first image which shows an extension agent in a field of decomposing radishes.  (See image)..

Google said this was "similar"  to the image I chose.

http://www.google.com/search?hl=en&imgrefurl=http://www.extension.org/pages/18524/how-cover-crops-suppress-weeds&imgurl=http://eorganic.info/sites/eorganic.info/files/u118/daikon-radish-cover.jpg&w=650&h=368&sig=117180872364197889543&ndsp=14&tbm=isch&tbs=simg:CAESEglr_1-MIUn3Z_1yEuEKhYqiNEvg&sa=X&ei=kcEdT_jWOO_CsQLcotGhCw&ved=0CAQQ0gU&biw=1276&bih=727

The images "similar" show everything from castles to turkey hunting to military personal walking through a jungle.

How in world does this make sense?

ergophobe

#1
January 23, 2012, 10:11:48 PM
I don't know about the Google algo, but at their simplest, image similarity algos will sample pxiel by pixel (either completely or randomly) and map the palette. So color is a big factor. Conversion to greyscale helps map shapes. I would expect Google to add in the face recognition they already own for Picasa where appropriate.

I'm thinking Google must have more sophisticated algos than that because that is presumably why they bought that visually-based comparison shopping engine some time back, but I expect that they all have to be based on some type of mapping of colors and greyscale values.

If you look at the images in your example, to me they are highly similar - the palettes match and they all have a lot of "noise" (not image noise, but just few large patches of the same color).

grnidone

#3
January 23, 2012, 10:29:40 PM Last Edit: January 23, 2012, 10:31:42 PM by grnidone
Quotef you look at the images in your example, to me they are highly similar - the palettes match and they all have a lot of "noise" (not image noise, but just few large patches of the same color).

Apparently, color is the *only* thing they are matching.  It is clear the words around the image have nothing to do with this particular algo.  And I don't understand why Google -- supposedly good at matching images to textual content -- isn't doing this.  

*wonders if JasonD has exploited this yet.*



dogboy

#4
January 23, 2012, 10:42:48 PM
I have seen some image recognition search engines.... best one was http://www.tineye.com/

ergophobe

#5
January 25, 2012, 12:14:22 AM
QuoteApparently, color is the *only* thing they are matching. 

Definitely also matching level of entropy. To me, at a quick glance, the images that come up in the two searches are really visually similar.

If you're working on a design and you're trying to find something visually similar but *not* of the same subject, it's more useful as is. I would say it depends on the goals before I could say one way or the other if the match is "good", but let's assume I'm from another planet and all I have in common with humans is that I see in the same visual spectrum, but I have no cultural context. In that case, I woudl say these images are highly similar.

ergophobe

#6
January 27, 2012, 12:29:38 AM
Interesting followup - a photographer friend is using this to find copyright violators.

dogboy

#7
January 27, 2012, 04:15:58 PM
If it isnt tineye, then its another similar engine that used by getty images to find their images and then they send snail mail notices to the whois with an outrageous formal looking bill that scares the shit out of people so they just pay it. The reality is if you just take down the image and do nothing, nothing happens.  But the letter is so formal people get fooled into thinking they have no choice.  They even send a snapshot of the image on your page.
Print
Go Up Pages1
User actions