You would need to provide at least 8 images for each model: Two different documents of the same type, for each of them, 2 front pictures and 2 back pictures.
Training Data Set Requirements
Please make sure that the training data set for a new document comply with the requirements below.
- 8 images are needed to train one document (with two sides):
- 2 images of each side of the same document (4 in total for two sides of the ID)
- 2 sets of images of the document from two different people (pictures of two physical copies of the same type of a document)
- Image quality requirements (if not met the training will not be successful)
- At least 3000*2000 pixels resolution
- The document should take up majority of the space on the image
- The document, or any of it's part is not covered by any object - paper, fingers ...
- The document is sharp, not blurred, no glares, all fields are clearly readable
- The whole document is visible on the image - all 4 corners of the document are visible
- Minimum height of the smallest letter is at least 12 px for each of the fields
- All fields are clearly legible by human (it is possible to distinguish even the most similar letters and diacritical marks without doubts)
- The document is pictured directly from above (or scanned on a high resolution scanner), not under any angle
- Physical requirements for the documents
The document is of a rectangular shape with sharp or only slightly round corners (think credit cards or passports)..
The size of the document does not exceed the size of an ISO 216 international standard paper size (A6, 105×148 mm), with the ratio of the sides ranging from 1:1 to 5:2.
The document is created from a solid material, or otherwise the users are instructed not to allow bending of the document while taking the image (e.g. by laying the document against a flat surface)
The character set of every field that should be processed is extended latin - including diacritics and common special characters.
The content of the fields is not moving around (filled in on a typewriter, custom printed on a standard office printer, glued in photographs etc)