Granted Patent B2

US 12,639,966 · App. 18/434,892 · Granted May 26, 2026

Inventors: Mridul Balaraman (Bangalore, IN); Madhusudan Singh (Bangalore, IN); Ashwin Kanth (Bangalore, IN); Pinak Pani Gogoi (Bangalore, IN); Sakshi Singh (Benares, IN)

Assignee: L&T TECHNOLOGY SERVICES LIMITED

G06V30/12G06F40/279G06V10/82G06V30/245

Abstract

Claims (53)

1 . A method of performing domain-based optical character recognition (OCR) error correction, the method comprising:

receiving, by a processor, at least one document image,

wherein the at least one document image comprises data corresponding to at least one domain;

determining, by the processor, one or more text entities in the data using an Optical Character Recognition (OCR) technique,

wherein the one or more text entities are determined based on recognition of each character of each of the one or more text entities;

for each of the one or more text entities:

determining, by the processor, a character embedding indicative of each character recognized in a corresponding text entity;

determining, by the processor, a style embedding indicative of one or more text-characteristic information of each character recognized in the corresponding text entity;

determining, by the processor, a layout embedding of each of the one or more text entities indicative of a layout hierarchy information of each of the one or more text entities in the at least one document image;

determining, by the processor, a concatenated embedding of each of the one or more text entities based on the corresponding character embedding, the style embedding and the document layout embedding of each of the one or more text entities; and

for each of the one or more text entities:

determining, by the processor, a corrected character embedding of each character recognized in the corresponding text entity based on the corresponding concatenated embedding using an encoder-decoder model,

wherein the encoder-decoder model is trained to recognize data corresponding to the at least one domain.

2 . The method of claim 1 , wherein the text-characteristic information for each of the text entities comprises font information of each character of each of the text entities,

wherein the font information comprises a font type, a font size, case information, and typography information.

3 . The method of claim 1 , wherein the layout hierarchy information comprises a plurality of hierarchy classifications comprising a table, a header, a sub-header, a paragraph, a graphical representation, etc.

4 . The method of claim 1 , wherein the concatenated embedding of each of the one or more text entities is determined by concatenating the corresponding character embedding, the style embedding, and the document layout embedding, respectively.

5 . A system of performing domain-based optical character recognition (OCR) error correction, the system comprising:

a processor; and

a memory communicably coupled to the processor, wherein the memory stores processor-executable instructions, which, on execution by the processor, cause the processor to:

receive at least one document image,

wherein the at least one document image comprises data corresponding to at least one domain;

determine one or more text entities in the data using an Optical Character Recognition (OCR) technique,

wherein the one or more text entities are determined based on recognition of each character of each of the one or more text entities;

for each of the one or more text entities:

determine a character embedding indicative of each character recognized in a corresponding text entity;

determine a style embedding indicative of one or more text-characteristic information of each character recognized in the corresponding text entity;

determine a layout embedding of each of the one or more text entities indicative of a layout hierarchy information of each of the one or more text entities in the at least one document image;

determine a concatenated embedding of each of the one or more text entities based on the corresponding character embedding, style embedding and document layout embedding of each of the one or more text entities; and

for each of the one or more text entities:

determine a corrected character embedding of each character recognized in the corresponding text entity based on the corresponding concatenated embedding using an encoder-decoder model,

wherein the encoder-decoder model is trained to recognize data corresponding to the at least one domain.

6 . The system of claim 5 , wherein the text-characteristic information for each of the text entities comprises font information of each character of each of the text entities,

wherein the font information comprises a font type, a font size, case information, and typography information.

7 . The system of claim 5 , wherein the layout hierarchy information comprises a plurality of hierarchy classifications comprising a table, a header, a sub-header, a paragraph, a graphical representation, etc.

8 . The system of claim 5 , wherein the concatenated embedding of each of the one or more text entities is determined by concatenating the corresponding character embedding, the style embedding, and the document layout embedding, respectively.

9 . A non-transitory computer-readable medium storing computer-executable instructions for performing domain-based optical character recognition (OCR) error correction, the computer-executable instructions configured for:

receiving at least one document image,

wherein the at least one document image comprises data corresponding to at least one domain;

determining one or more text entities in the data using an Optical Character Recognition (OCR) technique,

wherein the one or more text entities are determined based on recognition of each character of each of the one or more text entities;

for each of the one or more text entities:

determining a character embedding indicative of each character recognized in a corresponding text entity;

determining a style embedding indicative of one or more text-characteristic information of each character recognized in the corresponding text entity;

determining a layout embedding of each of the one or more text entities indicative of a layout hierarchy information of each of the one or more text entities in the at least one document image;

determining a concatenated embedding of each of the one or more text entities based on the corresponding character embedding, the style embedding and the document layout embedding of each of the one or more text entities; and

for each of the one or more text entities:

determining a corrected character embedding of each character recognized in the corresponding text entity based on the corresponding concatenated embedding using an encoder-decoder model,

wherein the encoder-decoder model is trained to recognize data corresponding to the at least one domain.

10 . The non-transitory computer-readable medium of claim 9 , wherein the text-characteristic information for each of the text entities comprises font information of each character of each of the text entities, and

wherein the font information comprises a font type, a font size, case information, and typography information.

11 . The non-transitory computer-readable medium of claim 9 , wherein the layout hierarchy information comprises a plurality of hierarchy classifications comprising a table, a header, a sub-header, a paragraph, a graphical representation, etc.

12 . The non-transitory computer-readable medium of claim 9 , wherein the concatenated embedding of each of the one or more text entities is determined by concatenating the corresponding character embedding, the style embedding, and the document layout embedding, respectively.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (15)