ML Table Extraction from Scanned Documents
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for extracting tables from scanned and image-based documents are inefficient due to errors introduced by optical character recognition (OCR) processing and require significant resource-intensive efforts, especially when dealing with degraded documents and multiple-page tables.
Innovation Solution
A machine learning-based system that identifies word bounding boxes, determines horizontal and vertical white space, and uses Natural Language Processing and machine learning algorithms to classify rows and extract table cell values, converting tables into computer-readable formats efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If Optical Character Recognition (OCR) is used to convert image-based documents into computer-readable characters, then text extraction is enabled, but accuracy deteriorates due to artifacts such as blurring and fading in degraded documents
Solution Approach 1:
The patent segments the document processing into distinct phases: first identifying table structures and boundaries, then extracting text within those structured regions. This segmentation allows the system to leverage structural information to improve character recognition accuracy in degraded areas, as table constraints provide contextual guidance for OCR in challenging regions.
Solution Approach 2:
The system performs preliminary table structure identification and cell boundary detection before conducting OCR. By establishing the table framework first, the system prepares contextual information that guides subsequent text extraction, improving accuracy by knowing the expected structure and content boundaries before attempting character recognition in potentially degraded areas.
2Loss of information
If traditional table extraction methods are used, then some table data can be retrieved, but efficiency deteriorates due to resource-intensive processing and manual effort requirements
Solution Approach 1:
The patent replaces manual mechanical table identification and text extraction processes with automated machine learning models. These models automatically detect table structures, identify cells, and extract content without human intervention, dramatically improving productivity while maintaining high data retrieval accuracy through learned patterns from training data.
Solution Approach 2:
The system implements self-service table extraction where the machine learning model autonomously performs all steps from table detection to content extraction. The model self-corrects and adapts through training, eliminating the need for manual configuration or intervention, thereby achieving both high information retrieval and computational efficiency.
3Measurement precision
If manual table identification and extraction methods are used, then accuracy can be maintained through human review, but productivity deteriorates due to significant manual effort and time requirements
Solution Approach 1:
The patent incorporates feedback mechanisms where the machine learning model's predictions are evaluated against ground truth or user corrections. This feedback loop continuously improves the model's accuracy over time, allowing the system to achieve human-level or superior accuracy automatically without requiring manual review for each document, thus maintaining precision while dramatically improving productivity.
Data Source
AI summary
In some embodiments, a method includes identifying a set of word bounding boxes in a first electronic document, and identifying locations of horizontal white space between two adjacent rows from a set of rows in a table. The method includes determining, using a Natural Language Processing algorithm, an entity name from a set of entity names for each table cell from a set of table cells in the table. The method includes determining, using a machine learning algorithm a class from a set of classes for each row from the set of rows. The method includes extracting a set of table cell values associated with the set of table cells, and generating a second electronic document including the set of table cell values arranged in the set of rows and the set of columns such that the set of words in the table are computer-readable in the second electronic document.