3

I'm using the concept of a perceptual hash for my images. It works well using the DCT as the underlying maths.

I get good matches for images which are the same bar some "attack" i.e. aspect ratio, brightness, contrast and most things related to color change.

However, I am looking for a bit more flexibility and want to be able to find images of similar nature, using an image to search other images. Similar to what TinyEye can do

The DCT is getting me the lowest frequencies of the image (structure). Is there something I could do here that can give me this added flexibility?

DaBler
  • 2,695
  • 2
  • 26
  • 46
Science_Fiction
  • 3,403
  • 23
  • 27
  • What about using your transformed data to compare the deltas between adjacent pixels? If the results are effectively the same, then wouldn't that handle an image which was the same but brighter? What if the deltas are different but still proportional to the source image? Wouldn't that be an image that is perhaps just a higher contrast version? – JDischler Oct 25 '12 at 18:07
  • So my current version works well for brightness, contrast and anything really to do with changing "color" on the pixels. I want flexibility in terms of if the image is taken from a slightly different angle. – Science_Fiction Oct 25 '12 at 18:17
  • Ah, apologies, I read this: "I get good matches...BAR some "attack" i.e. brightness" as "I get matches for images EXCEPT those with brightness, etc, differences" – JDischler Oct 25 '12 at 18:21
  • Slightly different angle implies a projective transformation, so you might be looking for projective invariants. The cross ratio of four features along a line might be one example you could use. It's tricky, though. – MvG Oct 26 '12 at 06:28
  • 1
    A classic Computer Vision approach to this would be interest point extraction (aka "features"), such as SIFT (patented and slow, but good), SURF (faster, patented as well) and a bunch of others. Feature matching gives a great deal of flexibility over a wide range of transformations. It involves a bunch of non-trivial and potentially expensive operations though, such as nearest-neighbor searches in high-dimensional spaces. – DCS Feb 07 '13 at 22:15

0 Answers0