Synthetic banknote data generation using a generative adversarial network with spatially composited multispectral data
A method for generating synthetic banknotes requires that a multispectral image be generated from a sample banknote. The multispectral image is processed to create a training image in a two-dimensional space. A generative adversarial network is trained using the training image. Synthetic banknotes are generated by seeding the trained generative adversarial network with random data. The synthetic banknotes may then be used to generate a banknote template for a currency validator.
1. A method for generating a set of synthetic banknotes, comprising:
generating a multispectral image of a sample banknote,
processing the multispectral image to create a training image in a two-dimensional space, wherein processing the multispectral image to create a training image in a two-dimensional space comprises:
generating a first false color image based on reflective infrared characteristics from the multispectral image,
generating an obverse color image from the multispectral image,
generating a reverse color image based from the multispectral image,
generating a second false color image based on transmissive infrared and transmissive green characteristics from the multispectral image, and
combining the first false color image, the obverse color image, the reverse color image, and the second false color image to be an interim version of the training image and then adding padding to the interim version of the training image so that the resultant training image has a square shape;
training a generative adversarial network using the training image; and
generating the set of synthetic banknotes by seeding the trained generative adversarial network with random data.
2. The method of claim 1 , wherein the generative adversarial network consists of a generator model and a discriminator model.
3. A method for generating a banknote template for a currency validator, comprising:
generating a multispectral image of a sample banknote,
processing the multispectral image to create a training image in a two-dimensional space, wherein processing the multispectral image to create a training image in a two-dimensional space comprises:
generating a first false color image based on reflective infrared characteristics from the multispectral image,
generating an obverse color image from the multispectral image,
generating a reverse color image based from the multispectral image,
generating a second false color image based on transmissive infrared and transmissive green characteristics from the multispectral image, and
combining the first false color image, the obverse color image, the reverse color image, and the second false color image to be an interim version of the training image and then adding padding to the interim version of the training image so that the resultant training image has a square shape;
training a generative adversarial network using the training image;
generating a set of synthetic banknotes by seeding the trained generative adversarial network with random data; and
generating the banknote template from the set of synthetic banknotes.
4. The method of claim 3 , wherein the generative adversarial network consists of a generator model and a discriminator model.
5. A method for generating a set of synthetic images, comprising:
generating a multispectral image of a sample image,
processing the multispectral image to create a training image in a two-dimensional space, wherein processing the multispectral image to create a training image in a two-dimensional space comprises:
generating a first false color image based on reflective infrared characteristics from the multispectral image,
generating an obverse color image from the multispectral image,
generating a reverse color image based from the multispectral image,
generating a second false color image based on transmissive infrared and transmissive green characteristics from the multispectral image, and
combining the first false color image, the obverse color image, the reverse color image, and the second false color image to be an interim version of the training image and then adding padding to the interim version of the training image so that the resultant training image has a square shape;
training a generative adversarial network using the training image; and
generating a set of synthetic images by seeding the trained generative adversarial network with random data.