Quote:
Originally Posted by Nathan
borked, very nice...
now, if we could only setup our own industry wide video fingerprinting so we do not have to pay some company to run it, that would be lovely... q is, can we without using some patent that possibly exists?
Your current hashes, they change if the image is resized (preserving aspect) or lowered in quality, right? So how can we build hashes which are still accurate enough but do not care about resizing or quality loss?
Any thoughts on that?
I am wondering if changing resolution of an image to a very low number, like 50x50 or so, if the colors would get close enough together regardless of how the image is cut or changed in quality?
IE, take a square part of the inside of an image of around 1000x1000, re-size it to 50x50 using a standard re-size technique which interpolates the colors.
Then use this on two versions of the same image, jpeg at 100 and jpeg at 50% quality.. and see what happens to the outcome, compare it visually...
|
Nathan - no, the hash db is constant as that is the reference data set. The hashes for the queries are created on the fly (these aren't one-way ie md5 hashes by the way but floating reference point hashes) and used to query the hash db.
See my point above for distorting images - I used the imagemagick command
mogrify -format jpg -define jpeg:size=240x180 -thumbnail 120x90 '*.jpg'
to make thumbs of all the images in the query set and all came back with the original image found in the db. 2 false negatives slipped through.
This is only for images, not videos and no patents were violated in its creation.