Searching for images | |||||||||||||||||
Another anti-patent posting. Build a distribution of spatial frequencies in an image, by energy density (self dot product) and normalize. This eliminates phase (direction) and brightness dependencies. Take a reference image at some scale and dot product this with the test image. The result will have a high value if the correlation is high. This is independent of both phase (relative orientation of the images) and position of the corresponding image fragment in the test image. Repeat with different scales of reference image. Better yet, fuzz the reference image so that the frequency distribution corresponds to several slightly differently scaled reference images. This will allow faster scanning at the expense of false positives, and will more likely catch perspective-distorted images. When a hit occurs, perform dot products with phase-preserved frequency distributions to weed out the false positives. This wil involve rotating and perspective-distorting the reference. When a hit passes that test, a more conventional scan of the image can be done now that the scale, orientation and perspective distortion are known. Finally, give it (and context information) to the Mk1 eyeball to counter optical illusions, etc. As an alternative, other phase-space transforms can be used instead of frequency distributions, an example being partial differential equations. Both test and reference images can be preprocessed, e.g. with edge enhancement or noise reduction. In most cases, the same preprocessing should be applied to test and reference.
nihil, Apr 11 2007
What do you think of this idea or comment? | |||||||||||||||||
Users who liked this idea also liked: | ||||||||||||||||||||||||||||||||
| ||||||||||||||||||||||||||||||||
Add your comment
This was way beyond my understanding, sorry. Um, what is the goal?
the only way this idea will ever be marketable is if you include radium in some way