Neural Network Document Extraction via OCR-Image Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data extraction systems from scanned documents are inefficient due to the need for manual labor, errors in optical character recognition (OCR), and the requirement for large amounts of training data for machine learning models, especially when dealing with documents of varying formats and layouts.
Innovation Solution
A system utilizing a neural network and optical character recognition (OCR) to identify regions of interest in scanned documents, such as tables, without relying on pre-defined rules or extensive training data, allowing for efficient extraction of relevant data from documents with diverse layouts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional machine learning techniques are used for document extraction, then the system can learn document layouts and information locations, but it requires thousands of example documents as training data, which increases costs and complexity
Solution Approach 1:
The system segments the document processing task into two independent components: (1) OCR-based text recognition that converts document images to machine-readable text with location information, and (2) rule-based extraction that uses regular expressions and patterns to extract specific information. This segmentation eliminates the need for large training datasets while maintaining extraction accuracy.
Solution Approach 2:
The patent replaces the mechanical learning process (neural networks requiring thousands of training examples) with a rule-based system using regular expressions and pattern matching. This substitution leverages the structured nature of document data to achieve accurate extraction without requiring the system to 'learn' from extensive training data.
2Device complexity
If zero shot learning is used to reduce training data, then the number of example documents decreases, but hardware costs increase and performance degrades with more fields
Solution Approach 1:
The system uses self-service through template-based extraction where pre-defined patterns and regular expressions automatically identify and extract information. The rules are designed to be field-agnostic, allowing the same extraction mechanism to handle any number of fields without performance degradation or additional hardware requirements.
Solution Approach 2:
The rule-based extraction system is universal and can handle any document field or format by adjusting the extraction rules rather than requiring retraining. A single system architecture serves multiple extraction needs across different document types and field configurations without additional hardware costs or performance degradation.
3Productivity
If OCR technology is used to automatically translate text data, then manual labor is reduced, but all data including irrelevant data is produced requiring human intervention to determine relevant data
Solution Approach 1:
The system extracts only the relevant information needed from the OCR output by using targeted regular expressions and extraction rules. Instead of processing all OCR data, the system selectively extracts specific fields (e.g., invoice numbers, dates, amounts) based on predefined patterns, automatically filtering out irrelevant data without human intervention.
4Extent of automation
If pre-defined rules are used to extract information, then extraction can be automated, but they cannot reliably extract information from invoices with different formats and layouts
Solution Approach 1:
The extraction rules are designed to be dynamic and adaptable to different document formats. The system uses flexible pattern matching and regular expressions that can adjust to varying layouts, positions, and formats of document elements. This allows the same automated extraction system to handle diverse invoice formats without requiring manual reconfiguration for each format.
Data Source
AI summary
A system to extract data from regions of interest on a document is provided. The system includes a storage device storing an image derived from a document having text information. The system includes a document importer operable to perform optical character recognition to convert image data in the image to machine readable data. The system includes a neural network that identifies at least one region of interest on the image to classify an area of the at least one region of interest as a table. The neural network is operable to take as input the machine readable data and the image and combine both the machine readable data and the image to determine that the classified area is the table.