Finding duplicated photo

Billy Mays noway at nohow.com
Fri Jul 8 10:32:13 EDT 2011


On 07/08/2011 10:14 AM, TheSaint wrote:
> Billy Mays wrote:
>
>> It worked surprisingly well even
>> with just the 64bit hash it produces.
>>
> I'd say that comparing 2 images reduced upto 32x32 bit seems too little to
> find if one of the 2 portrait has a smile referred to the other.
> I think it's about that mine and your suggestion are similar, but I'd like
> to scale pictures not less than 256x256 pixel.
> Also to take a wider case which the comparison involve a rotated image.
>

Originally I thought the same thing.  It turns out that doing a DCT on 
an image typically moves the more important data to the top left corner 
of the output.  This means that most of the other data in the output an 
be thrown away since most of it doesn't significantly affect the image. 
  The 32x32 is an arbitrary size, you can make it any square block that 
you want.

Rotation is harder to find.  You can always take a brute force approach 
by simply rotating the image a couple of times and try running the 
algorithm on each of the rotated pics.  Image matching is a difficult 
problem.

--
Bill



More information about the Python-list mailing list