Computer: Searching: Image
Search for Images Using an Image   (+14, -2)  [vote for, against]
Wouldn't it be nice.

Have you ever searched for an image, found it, but the link was broken? You had no formal name for the image, the description wouldn't bring up any other links, but you had the thumbnail image with no corresponding full size image. You can usually find what you want after enough text searching and browsing, but it'd be bitchin If you could use that thumbnail to search for the larger image. I don't know if this is technically possible, but would be useful.
-- miggavin, Apr 01 2006

CIRES http://amazon.ece.u...~qasim/research.htm
Content-Based Image Retrieval System [jutta, Apr 01 2006]

Only if you can draw really well cartoogle
Proof of concept [methinksnot, Apr 01 2006]

GooglePix GooglePix_20search
[DesertFox, Apr 01 2006]

Retrievr - Flickr
[Dub, Apr 03 2006]

http://labs.systemo...S9pbWcvaGMuZ2lm.gif Images similar to a croissant on a plate. [spidermother, Sep 05 2009]

[Dub, Sep 05 2009]

The boffins at Google have been through the 'bakery, for sure... [neutrinos_shadow, Sep 23 2010]

This is like the porn facial recognition search in this category, but widely applied.
-- miggavin, Apr 01 2006

Yeah jutta, that's good, but you can't use your provided pic to search.
-- miggavin, Apr 01 2006


I think that isn't the right adjective to use here. Also, your first sentence needs a question mark.
-- sninctown, Apr 01 2006

I like it
-- silentman, Apr 02 2006

How is a search engine going to know what it is about that image that you are searching for?
-- Galbinus_Caeli, Apr 03 2006

Since in [miggavin]'s case we have a thumbnail that's mechanically related to the sought-after image, you don't need to answer that.

In the more general case, at worst, you'd perform scene analysis - simulate whatever it is a person does when you show them an image and they figure out what it "means". Not easy, not impossible, existing Computer Vision research topic.
-- jutta, Apr 03 2006

[jutta] That is kind of my point. Computer vision is crap at the moment. It is supposed to get better in the next five years, and has been that way for fourty years. It is really easy for a human to tell that a picture represents a croissant on a plate rather than nude model on a bed, or a turd in a puddle, but grapically these are very similar things.
-- Galbinus_Caeli, Apr 03 2006

But, Galbinus, miggavin isn't searching for *similar* images, s/he is "bitchin after" other copies of the *same* image (for the metadata), or scaled-up versions thereof.

If thumbnails were all the same size, didn't undergo alternative lossy compressions, and didn't suffer from bit-rot (re-encoding with a lossy format), then it would be relatively easy. Find all images, generate their thumbnails, hash them.
But they are and do, so it is a bit more tricky.

Perhaps wavelet-style compression would be useful?
Divide the image into quadrants, work out the average colour for each, and store it. Iterate one or more times on the quads for increasing accuracy. This could give quite an efficient multi-levelled seaching algorithm.
It also has the advantage of coping with distorted aspect ratios, but wouldn't necessarily find cropped or colour-processed derivatives*. You may be able to deal with the latter to some extent by pre-processing to get an average colour balance, but for the former... you can't have everything.

*In the thumbnail case it itself would be the derivative, but I'm thinking it would be a useful plagarism-detection tool as well.
-- Loris, Apr 03 2006

Thumbnail to larger version might be possible, if the algorithm for the thumbnail generation were known. But if the source image had been processed further than it would be very difficult to identify.
-- Galbinus_Caeli, Apr 03 2006

Baked - Like this? (link)
-- Dub, Apr 03 2006

I would not call this baked as described. [Dub]'s link finds somewhat similar images, but not the _same_ image. It also returns only Flickr images.

A croissant on a plate (link) returns mainly dogs; no nude model on a bed or turd in a puddle :-(.
-- spidermother, Sep 05 2009

-- Dub, Sep 05 2009

I was just about to post the idea below when I found this "Search for Images using an Image" while looking for a category to put the idea in. The links listed by readers are great, as are the links that the links lead to. - Thank you!

Often I want a part to build something with. Because there is hardly anything new under the sun, and I don't want to spend the time and expense fully designing designing and building it, I would like to make a rough drawing or CAD drawing of the part and submit it to an image search engine that will use image recognition to find similar parts on the Internet.
-- Sunstone, Sep 22 2010

[Sunstone] You should spin that off as a separate idea. You really want to match on product dimensions, etc., not mere appearance, and the manufacturer shirley has that info in machine-readable format. If CAD-CAM file formats are standard enough, manufacturers could expose their products' specs to the internet well enough to permit a match, without revealing copyright or proprietary information: Google already does this with a lot of scholarly books & papers, where you get a link to something that matches your search terms, and if you click on it, you get a paywall login screen.
-- mouseposture, Sep 23 2010

This, and Goggles will be fantastic when available across a few more platforms.
I imagine that an image/ image-text based search engine will spawn all sorts of new devious practices, such as painting your building to look like someone else's building, and web sites will need to have photos of the outside views of their premises.
-- Ling, Sep 24 2010

It's a shame that not all images are 2 to the power of something in dimension. If they were and were also encoded in losless JPG, then there would be no need to reconstruct the image as the the thumbnail would be the first set of coefficients of the larger image and thus easy to compare. Still possible but more complicated on non power of 2 numbers.

A good starting point for registering images would be to construct a table of the average colour and brightness of the image as well as the average and standard deviation of the gradient of each image. This would allow the search to be narrowed down before soing anything more complicated.

Though, having said all that, when skynet comes online we can just ask it for answers...
-- saedi, Sep 24 2010

We won't need to ask. It'll just tell us what it feels we should know.
-- mouseposture, Sep 25 2010

bun for welcoming our new Skynet overlord.
-- Voice, Sep 25 2010

I need this!
-- Jscotty, Feb 22 2011

Google goggles, as linked! I found the name of a print I've had on the wall for years with it.
-- Zimmy, Feb 23 2011

random, halfbakery