Image Search

[shaunbhoy]

Good call. But to be honest I was also hoping I would get some sort of proximity match as well.

Not sure of the site, but I remember one where you would put a picture of yourself and then it would come back with a whole list of people on the internet who looked like you.

I always get Brad Pitt first.

Don't want to rain on your parade MF, but everyone gets that. It is just an image they put up whilst the background search carries on. Admit it, after that you got...........in the following order..........

1) Deputy Dawg
2) Dobbie
3) Cartman
4) Jimmy Krankie
5) John Candy
6) Marc Almond
7) Jabba the Hutt
8) Mr Tumnus
9) John Belushi
10) Jar Jar Binks

[MarillionFan]

Good call. But to be honest I was also hoping I would get some sort of proximity match as well.

Not sure of the site, but I remember one where you would put a picture of yourself and then it would come back with a whole list of people on the internet who looked like you.

I always get Brad Pitt first.

[TimberWolf]

This thread on Reddit suggests great cleverness in the use of Fourier Transforms. My initial guess was Discrete Cosine Transforms, but there are comments on there explaining why they can't work (or not in any usable way) for partial image matching.

I ended up feeling a bit stupid for having thought of DCTs in the first place when they obviously can't work. I'd underestimated the limitations imposed by the variance under transformation (amongst other things), but I was slightly reassured by the fact that I may, if I think very hard for a very long time, be able to work out all the stuff suggested there about FTs

Footnote: I have a copy of Byte from around 1984 that proposes the Fast Hartley Transform as a more computationally efficient alternative to the Fast Fourier Transform. I assume this is of limited value now, when anybody who needs efficient FFT processing can just buy hardware that does it for pennies. It reminds me of all the time I spent on Bresenham line-drawing algorithms (also known as digital differential analysers) - time wasted once it became cost-effective to draw lines in hardware. Ho hum

It's a fascinating subject, but alas as with so many things, is probably too big a one to get engrossed in. Not least on how animal life handles image recognition.

I had a quick play with the Leon image, out of curiosity. The image was still recognised by TinEye following a 10 degree rotation, but not by a 30, 90 or 180 degree rotation.

I once put some basic image recognition into a piece of furniture making software I wrote, which worked well despite my lack of research into available algorithms. It just had to identify and categorise furniture items. Fortunately I was working with image profiles (skeleton, line images) which I reduced to a (IIRC) a 5*5 block grid that was used to compare against stored templates held in a similar format. It handled small rotations too.

[NickFitz]

+1 from me too. That worked better than I'd expected. I wish I was clever enough to know how they did it.

This thread on Reddit suggests great cleverness in the use of Fourier Transforms. My initial guess was Discrete Cosine Transforms, but there are comments on there explaining why they can't work (or not in any usable way) for partial image matching.

I ended up feeling a bit stupid for having thought of DCTs in the first place when they obviously can't work. I'd underestimated the limitations imposed by the variance under transformation (amongst other things), but I was slightly reassured by the fact that I may, if I think very hard for a very long time, be able to work out all the stuff suggested there about FTs

Footnote: I have a copy of Byte from around 1984 that proposes the Fast Hartley Transform as a more computationally efficient alternative to the Fast Fourier Transform. I assume this is of limited value now, when anybody who needs efficient FFT processing can just buy hardware that does it for pennies. It reminds me of all the time I spent on Bresenham line-drawing algorithms (also known as digital differential analysers) - time wasted once it became cost-effective to draw lines in hardware. Ho hum

[PRC1964]

+1 from me too. That worked better than I'd expected. I wish I was clever enough to know how they did it.

[NickFitz]

TinEye is great for image searches. I haven't tried searching on partial images though.

+1. It seems to work pretty well for partial matches: for example, searching for Platypus's avatar returned this among its 400+ results:

[TimberWolf]

TinEye is great for image searches. I haven't tried searching on partial images though.

[Paddy]

Check out No.7

That link has some nasty scripts running!

[Clippy]

Check out No.7

[Platypus]

Anyone know of a search engine where you can give it a partial image and it will search for the full size image?

Serious question. Although I expect some dodgy answers