Granted Patent B2

US 12,444,163 · App. 18/419,946 · Granted Oct 14, 2025

Inventors: Mehrdad Jabbarzadeh Gangeh (Mountain View, CA); Hamid Reza Motahari Nezad (Los Altos, CA)

Assignee: EYGS LLP

G06V10/454G06F16/258G06N3/045G06N3/088G06V10/82G06V30/19173G06V30/412G06V30/414

Abstract

Claims (39)

1. A non-transitory, processor-readable medium comprising code that when executed cause a processor to:

remove, using a mask, at least some lines from a first set of tables to generate a second set of tables;

train a conditional generative adversarial network (cGAN), using the first set of tables and the second set of tables, to produce a trained cGAN, the cGAN including a generative algorithm and a discriminative algorithm, the discriminative algorithm configured to authenticate table pairs from the first set of tables and the second set of tables based on a predicted table produced by the generative algorithm, each of the generative algorithm and the discriminative algorithm including at least one convolutional neural network layer;

identify, using the trained cGAN, a plurality of lines for overlaying onto a lineless table; and

overlay the plurality of lines onto the lineless table to produce a lined table.

2. The non-transitory, processor-readable medium of claim 1 , wherein the first set of tables includes a plurality of table images.

3. The non-transitory, processor-readable medium of claim 1 , wherein the first set of tables includes a plurality of table images, and the non-transitory, processor-readable medium further comprises code that when executed cause the processor to:

convert the plurality of table images into a binary format based on an adaptive threshold.

4. The non-transitory, processor-readable medium of claim 1 , wherein the cGAN is a supervised GAN.

5. The non-transitory, processor-readable medium of claim 1 , wherein the first set of tables and the second set of tables collectively includes tables with lines and tables without lines.

6. The non-transitory, processor-readable medium of claim 1 , wherein:

each table in the first set of tables includes lines, and

each table in the second set of tables does not include lines.

7. The non-transitory, processor-readable medium of claim 1 , wherein the cGAN includes an atrous convolutional neural network layer.

8. A non-transitory, processor-readable medium comprising code that when executed cause a processor to:

generate training data, via a processor, the training data including:

a first dataset including at least one of lineless formatted data or partially lined formatted data, and

a second dataset including formatted data with format lines, at least some of the format lines from the second dataset removed using a mask to generate the first dataset; and

train an artificial neural network (ANN) using the training data based on a local visual structure of the training data and a global visual structure of the training data, to produce a trained ANN configured to predict line placement for at least one of a lineless table or a partially lined table, the ANN including a generative algorithm and a discriminative algorithm, the discriminative algorithm configured to authenticate table pairs from the first dataset and the second dataset based on a predicted table produced by the generative algorithm, each of the generative algorithm and the discriminative algorithm including at least one convolutional neural network layer.

9. The non-transitory, processor-readable medium of claim 8 , wherein the ANN is a conditional generative adversarial network (cGAN).

10. The non-transitory, processor-readable medium of claim 8 , further comprising code that when executed cause the processor to:

calculate a quality metric via the discriminative algorithm included in the ANN,

the code to train including code to train the ANN based on the quality metric.

11. The non-transitory, processor-readable medium of claim 8 , wherein:

the ANN is a conditional generative adversarial network (cGAN), and

the cGAN is a supervised GAN.

12. The non-transitory, processor-readable medium of claim 8 , further comprising code that when executed cause the processor to:

overlay at least one line at the at least one of the lineless table or the partially lined tables based on predicted line placement generated by the trained ANN.

13. The non-transitory, processor-readable medium of claim 8 , wherein the trained ANN includes an atrous convolutional neural network layer.

14. A processor-implemented method, comprising:

receiving a first table that is one of a lineless table or a partially lined table; and

generating a second table that is a lined table based on the first table and using a trained neural network model that predicts line placement based on structural attributes of the first table, the trained neural network model trained using tables with lines and tables without lines, at least some lines from the tables with lines removed using a mask to generate the tables without lines, the trained neural network model including an atrous convolutional neural network layer, the structural attributes of the first table including at least one of global visual structural attributes of the first table or local visual structural attributes of the first table.

15. The processor-implemented method of claim 14 , wherein the trained neural network model is a cGAN.

16. The processor-implemented method of claim 14 , wherein the generating the second table includes determining at least one predicted line and superimposing the at least one predicted line onto an image of the first table.

17. The processor-implemented method of claim 14 , further comprising:

sending a signal to cause display of the second table within a graphical user interface (GUI).

18. The processor-implemented method of claim 14 , wherein:

the trained neural network model is a cGAN that includes at least one convolutional neural network layer, the atrous convolutional neural network layer included in the at least one convolutional neural network layer.

19. The processor-implemented method of claim 14 , wherein the trained neural network model is a cGAN that includes a generative algorithm and a discriminative algorithm, the discriminative algorithm configured to authenticate table pairs from a first set of tables and a second set of tables based on a predicted table produced by the generative algorithm.

Assignments (2)

Continuity (2)

References Cited (178)