The so-called "Normalized Compression Distance" measure (NCD) can be used for this kind of thing. I used it for OCR of handwritten digits in the paper "Clustering by Compression". You will probably have to fiddle with a small script to translate each image into a new file that is uncompressed and probably quantized in both pixelspace and colorspace. You can group similar images using "maketree" but it will only do about 200 images at a time maximum before it is too slow on most computers. Just check out the libcomplearn and libqsearch packages to find out more, or have a look on [1].
Best regards, Rudi [1]: http://complearn.org/ On Feb 6, 2008 6:11 PM, Paul E Condon <[EMAIL PROTECTED]> wrote: > I have a large number of scanned images of photographs. They are > scanned from prints, and for some images there are several prints. I > would like to find a program that would group similar images (similar > in some sense that approximates human perception of similarity) and > flags the slight differences within a group. I've looked on Google and > found a few programs for criminal investigation but they want $ and > seem only to exist for Windows, which I don't have here. > > I imagine that similarities could be found by image subtraction with > minimization of some measure of image difference. But I also imagine > that there are a host of complications. > > Suggestions? > -- > Paul E Condon > [EMAIL PROTECTED] > > > -- > To UNSUBSCRIBE, email to [EMAIL PROTECTED] > with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED] > > -- Which is worse, ignorance or apathy? Who knows? Who cares? --Erich Schubert -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]