5
$\begingroup$

I have an algorithm to to detect copy-move forgeries in images. I used block matching to detect regions of an image that were forged with copy-move forgery and highlighted the alleged areas. For block matching, I took the DCT of each block and normalized with an extended 16x16 quantization matrix.

The problem:

  • this approach does not work if the copied region is rotated or scaled.

My question:

  • What transformation/similarity metric would be invariant to rotations, shifts, and scaling? (Alternatively, is there another method available for solving this problem?)

  • Additional points if the method is already available in MATLAB.

Thanks for the help.

The code for this project is available here: copy-move forgery detection.

Edit: I am aware of the Radial Gradient Transform, but I don't think it is invariant to all three.

Improve this question
$\endgroup$
  • 1
    $\begingroup$ there's wavelet families whose bases in essence contain mother wavelets and scaled versions; that would possibly solve the scale-invariance issue. However, this sounds like you'd have to come up with your "own" wavelet. $\endgroup$ – Marcus Müller Jun 16 '16 at 20:19
  • $\begingroup$ I'll have to consider wavelets more closely. Thanks for the input! $\endgroup$ – Brady Sheehan Jun 16 '16 at 21:49
  • 1
    $\begingroup$ Your question has a close 1D version here dsp.stackexchange.com/questions/25926/… $\endgroup$ – Laurent Duval Jun 16 '16 at 23:38
1
$\begingroup$

Use a methods called Error Level Analysis (ELA). It can solve your problem as many forensic applications use it to detected doctored reigns in digital images.

Improve this answer
$\endgroup$
1
$\begingroup$

There are several feature descriptors for an image that are scale, rotation and shift-invariant. Two examples are SIFT and SURF.

You can extract feature points from one images and see, if you find the same feature descriptors in another image. This will very likely identify a match between both images.

Improve this answer
$\endgroup$
  • $\begingroup$ Coming back a little late, can you detail why you term them "scale, rotation and shift-invariant"? $\endgroup$ – Laurent Duval Mar 11 '18 at 18:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.