Generate Forged Signature From Real one for data augmentation

[dl_geek_007]

i am contacting you Geeks hoping that you can help me with your opinion about a use case,

i am in internship period and i’m working on a signature verification problem,

this process needs real and forged signature,

and all i have is real signature (like 30 signature per person), and i have to generate fake signature (forged) from those real one in order to augment data and balance it for the main purpose ,which is signature verification,

and during my search i found amazing videos explaining GANS and conditional gans , pix2pix …

and i’m wondering if you have an idea how to fit this model or an other approach to generate forged signature close to the real one but not as match as real ,

her is an example to clear the view a little bit,

those are the 3 images real for person 8 (img1,9,17)

and those are the forged one :

those examples are token from cedar dataset,

hope you can help with that ,
and keep up the good work geeks,

[kenfus]

Are you sure that generating fake signatures by a GAN would match fake signatures generated by humans, especially “forging” experts?

You could try to retrain the last layer of a pretrained CNN and add a layer at the beginning, which accepts two pictures: the original signature and the signature, which should be checked if it is forged or not and then train it on forged / non forged pictures.

If you really want to train a GAN, there are many good videos and guides, just google it.