1

I'm working on a project about offline signature verification and I've tried SIFT/SURF algorithms (OpenCV) for comparisson of 2 signature images.

What I've noticed is that when I pass in 2 same pictures I get ~1000 keypoints but when I pass 2 pics of different signatures of same person I get just ~70-80. And when one of the passed pics is a signature of a different person but which has alike style I get ~50-60 keypoints. Some of the points also weren't matching each other at all like they were from 2 different locations.

It's clear to me that these algorithms aren't good for my task but I don't quite understand why.

Could anyone exaplin the reason to me from the maths/algo point of view?

Mary
  • 45
  • 8
  • 2
    SIFT is often used as a first step of image orientation, so finding and describing keypoints and matching them using various methods. When you are trying to find similar signatures maybe **template matching** is more appropriate, but I would not completely exclude the use of a feature matching algorithm. – Grillteller Nov 05 '19 at 07:48

1 Answers1

0

Signature verification is a very difficult task, lots of research efforts have been made but still they are not much accurate in comparing signature pairs

SIFT/SURF algorithms wouldn't be helpful here because model needs to learn a more complex set of features in order to compare signatures

There are some Deep learning based Offline signature verification models that you can see

Atinesh
  • 1,790
  • 9
  • 36
  • 57
  • Thank you! So SIFT/SURF are usually used for less complex tasks? Like more basic images comparisson/search – Mary Sep 16 '20 at 06:58
  • @Mary Yes, We can say that – Atinesh Sep 16 '20 at 12:23
  • @Mary You can also approach the problem in a traditional non-deep learning way, by manually extracting the complex set of features such as Geometric features, Mathematical transformations features, Texture features, Graphometric features, Directional features, Wavelet features, Shadow features, etc. I haven't tried non deep learning based methods yet, but I can give you some comprehensive list of papers if you need – Atinesh Sep 16 '20 at 12:32
  • I'd be really glad if you could share (if it's not hard of course). Could also be useful for others :) – Mary Sep 16 '20 at 13:29
  • @Mary Give me your email address then I will send its difficult to send it on SO comment – Atinesh Sep 17 '20 at 06:06