Document authenticity determination

1. A method for assessing if characters in a sample image are formed from a predefined font comprising:

forming a first embedded space representation for characters of the predefined font;

extracting sample characters from the sample image;

forming a second embedded space representation of the sample characters; and

comparing the first and second embedded space representations to assess if the sample characters are of the predefined font, wherein the first embedded space representation for the predefined font and the second embedded space representation of the sample characters, is a transformed representation in a mathematical space that has a lower number of dimensions than the number of dimensions in a normal space needed to fully characterize the untransformed sample characters or the untransformed predefined font;

wherein each of the embedded space representations is a fixed-length vector representation of the sample image or the predefined font.

2. The method of claim 1 , wherein forming a first embedded space representation for the predefined font comprises:

extracting training characters from at least one image, the training characters comprising the characters of the predefined font; and

forming a first embedded space representation for the predefined font using the training characters.

3. The method of claim 2 , wherein the training characters comprise characters from at least two fonts, one of which is the predetermined font.

4. The method of claim 2 wherein the at least one image comprises at least one image of at least one authentic document, and wherein the sample image is an image of a sample document to be tested for authenticity.

5. The method of claim 2 , wherein extracting training characters comprises, for each image of the at least one image, segmentation of the image to extract at least one text field.

6. The method of claim 1 , wherein extracting sample characters comprises segmentation of the sample image to extract at least one text field.

7. The method of claim 5 , wherein segmentation of the image to extract at least one text field comprises extracting a text field image of the at least one text field, and

segmentation of the text field image to extract character images.

8. The method of claim 7 , wherein segmentation of the text field image to extract character images comprises:

converting the text field image to grayscale;

thresholding the grayscale image of the text field using a threshold value;

removing all pixel groups in the threshold image that contain less than a number of pixels;

identifying all remaining pixel groups;

identifying the minimal bounding box for identified pixel groups; and

using the minimal bounding box to extract character images from the text field image.

9. The method of claim 7 , wherein segmentation of the text field image to extract character images comprises using character classification to:

obtain at least one character label; and

order character images into an ordered list of character images,

wherein using character classification comprises using a convolutional neural network to extract features from the characters images.

10. The method of claim 9 , wherein the convolutional neural network is trained using generated images, wherein the generated images are generated by:

randomly selecting a background and randomly selecting a font;

generating a test image comprising a character of the selected font on the selected background;

perturbing the test image by a random image processing process.

11. The method of claim 9 , wherein the ordered list of character images and the at least one character label are used to form a list of pairs, each pair being two images of the same character in either a font that is the same, or a font that is different.

12. The method of claim 11 , wherein a transforming function is formed using the list of pairs that is configured to convert input image into a real-valued vector, and

wherein each embedded space representation is formed by passing the predefined font or the sample image to the transforming function.

13. The method of claim 12 , wherein the transforming function is formed using a Siamese network.

14. The method of claim 1 , wherein the first embedded space representation is obtained from a reference database; and wherein the steps of extracting, forming the second embedded space representation, and comparing are repeated for every sample image that is processed to form at least one similarity score.

15. The method of claim 1 , wherein comparing the first and second embedded space representations comprises:

computing a plurality of similarity scores of the first and second embedded space representations for the sample characters; and

calculating a final similarity score by averaging the plurality of similarity scores.

16. A non-transitory computer-readable storage medium comprising executable instructions for performing the method of claim 1 .

17. The method of claim 12 , wherein the transforming function is configured to convert an input image into a real-valued vector of 2048 length.

18. The method of claim 12 , wherein the ordered list of character images is ordered alphabetically.