Automatic generation of training data for hand-printed text recognition

A non-transitory computer readable medium (CRM) storing computer readable program code for generating training data for hand-printed text recognition, the computer readable program code causes a computer to:

• obtain a structured document comprising:
• a document page image that includes text characters, and
• document metadata that associates each of the text characters to a document character label;
• obtain, from a database, a set of hand-printed character images and database metadata, wherein the database metadata associates each of the set of hand-printed character images to a database character label;
• generate a modified document page image by iteratively processing each of the text characters, wherein the iterative processing comprises:
• determining, based on a predefined condition, whether an individual text character should be replaced;
• selecting a replacement hand-printed character image from the set of hand-printed character images, wherein the selecting is based on matching the document character label of the individual text character to the database character label of the replacement hand-printed character image;
• scaling the replacement hand-printed character image to match a size of the individual text character in the document page image;
• and
• inserting the replacement hand-printed character image into the modified document page image, at a location based on a position of the individual text character, wherein
• the selecting, scaling, and inserting are each performed in response to determining that the individual text character should be replaced;
• and
• output a training file comprising:
• the modified document page image, and
• modified document metadata that associates each of a set of inserted hand-printed character images to an inserted character label.

A system for generating training data for hand-printed text recognition, the system comprising:

• a memory;
• a processor coupled to the memory, wherein the processor:
• obtains a structured document comprising:
• a document page image that includes text characters, and
• document metadata that associates each of the text characters to a document character label;
• obtains, from a database, a set of hand-printed character images and database metadata, wherein the database metadata associates each of the set of hand-printed character images to a database character label;
• generates a modified document page image by iteratively processing each of the text characters, wherein the iterative processing comprises:
• determining, based on a predefined condition, whether an individual text character should be replaced;
• selecting a replacement hand-printed character image from the set of hand-printed character images, wherein the selecting is based on matching the document character label of the individual text character to the database character label of the replacement hand-printed character image;
• scaling the replacement hand-printed character image to match a size of the individual text character in the document page image;
• and
• inserting the replacement hand-printed character image into the modified document page image, at a location based on a position of the individual text character, wherein
• the selecting, scaling, and inserting are each performed in response to determining that the individual text character should be replaced;
• and
• outputs a training file comprising:
• the modified document page image, and
• modified document metadata that associates each of a set of inserted hand-printed character images to an inserted character label.