US 12,718,607 B2Granted

System and method for identifying the educational content using artificial intelligence

AssigneeExtramarks Education India Pvt Ltd.

InventorsRitvik Kulshrestha, Tanay Karve, Deep Dwivedi, Abhra Das, Suman Gadhawal, Vipin Tripathi, Gaurav Sharma

Life of this patent

Estimating the term…

Loading patent…

Abstract

The present disclosure provides a system and a method for authoring and automatically digitizing the printed and well as handwritten educational content inside an input image with a plurality of input characters. The method comprises the steps of receiving and processing an input image by a pre-processing module to identify and obtain a plurality of input characters from the input image, digitizing the plurality of input characters into the output string representations by a digitizing module, combining the plurality of output string representations using a post-processing module which is outputted in a computer-readable format, delineative text, and a markup language. Further, the system is based on the Convolution Neural Network (CNN) that may provide an embedded editing module allowing one or more users to make corrections and create new content. The CNN system may be retrained iteratively based on the editing performed by the users on the output strings representations to ensure high accuracy, better performance, and to avoid/reduce the error rate.

Claims12

Dependent claims

A system for recognizing one or more handwritten characters, wherein the system comprising:

camera or scanner,
obtaining an input image including a plurality of input characters;
a pre-processing module operable to identify the received input image and further extract a plurality of input characters from the input image;
a digitizing module operatively connected to the pre-processing module, wherein the digitizing module being operable to digitize the plurality of input characters into a plurality of output string representations;
a post-processing module connected to the digitizing module, wherein the post-processing module being operable to combine the plurality of output string representations such as a mathematical expression, a chemical formula, a complex chemical structure, or a physics circuit diagrams;
an embedded editing module connected to the post-processor module of the system, wherein the embedded editing module being operable to allow one or more users to make corrections in the plurality of output string representations and further create new content;
and
wherein the system is formed on a Convolutional Neural Network (CNN) that is retrained iteratively based on the editing performed by the one or more users on the output string representations, wherein the iterative retraining continues for one in 2000 instances or less than or equal to a range of 0.01 percent to 0.05 percent until the incremental increase in the performance and reduction in the error rate;
wherein the attention mechanism comprising:
generating the output string representations by mapping LaTeX (Lambert's Tex) sequences to one or more regions of the input image, wherein the regions are derived by generating a feature map from the input image that is passed to an autoencoder to predict the sequence of LaTeX tokens.

7 more elements

A method for recognizing one or more handwritten characters, wherein the method comprising:

sequential receiving of an input image including a plurality of input characters;
extracting a plurality of input characters from the input image by a pre-processing module;
digitizing a plurality of input characters into a plurality of output string representations, by a digitizing module;
combining the plurality of output string representations by a post-processing module, wherein the plurality of output string representations refers to a mathematical expression, a chemical formula, a complex chemical structure, or a physics circuit diagram s;
allowing one or more users to correct and further create new content in the plurality of output string representations using an embedded editing module;
and
wherein the method is formed on a Convolutional Neural Network (CNN) that is retrained iteratively based on the editing performed by the one or more users on the output string representations, wherein the iterative retraining continues for one in 2000 instances or less than or equal to a range of 0.01 percent to 0.05 percent until the incremental increase in the performance and reduction in the error rate;
wherein the attention mechanism comprising:
generating the output string representations by mapping LaTeX (Lambert's Tex) sequences to one or more regions of the input image, wherein the regions are derived by generating a feature map from the input image that is passed to an autoencoder to predict the sequence of LaTeX tokens.

6 more elements

Dependencies are inferred from the claim text.