Document Assembly Using Layout Analysis and Invariant Fields
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for processing multipage electronic documents lack efficiency in automatically separating individual documents from multipage files without separators, often requiring mechanical handling and barcode detection, which can lead to incorrect assembly.
Innovation Solution
A method that utilizes training data reflecting document layouts to automatically split individual documents from multipage files by matching pages to predefined layouts and using invariant fields to identify and assemble documents, even in the presence of attachments, without the need for special separators.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separator sheets with barcodes are inserted between documents prior to scanning, then document separation can be detected, but mechanical handling complexity and labor increase significantly
Solution Approach 1:
The invention extracts and removes the separator sheet requirement from the document assembly process. Instead of relying on physical separators with barcodes, the system uses layout analysis and training data to identify document boundaries directly from the scanned pages, eliminating the need for mechanical separator handling while maintaining accurate document separation.
Solution Approach 2:
The invention replaces the mechanical separator-based system with an automated image processing and layout analysis system. By using training data from document layouts and invariant field detection, the system substitutes physical separators and barcode scanning with computational methods that automatically identify document boundaries.
2Reliability
If barcode detection is used to identify separator images, then document boundaries can be detected, but assembly errors occur when barcodes are not read correctly
Solution Approach 1:
The invention introduces layout analysis and training data as an intermediary between the scanned pages and document boundary detection. Instead of directly relying on barcode reading, the system uses learned document structures and invariant fields to identify boundaries, providing a more robust detection mechanism that doesn't fail when barcodes are damaged or misread.
Solution Approach 2:
The system performs preliminary training by collecting layout data from sample documents to establish expected document structures and invariant fields. This preliminary action creates a reference model that guides the document assembly process, allowing the system to identify boundaries based on structural patterns rather than relying solely on potentially error-prone barcode detection.
3Extent of automation
If deep learning models or Markov models are trained to separate documents, then automation can be achieved, but the approach lacks utilization of layout information and training data
Solution Approach 1:
The invention applies local quality by focusing on specific invariant fields and layout characteristics that are unique to each document type and originator. Instead of using generic deep learning models, the system tailors the analysis to local document structures such as invoice numbers, dates, and formatting patterns, significantly improving assembly accuracy for specific document types.
Solution Approach 2:
The system performs preliminary training by collecting and analyzing layout data from sample documents of each originator. This creates originator-specific training data that captures the unique characteristics of each vendor's document format. This preliminary action enables the automated system to adapt to different document styles and improve reliability for each specific originator.
Data Source
AI summary
An automated method for assembling common commercial documents such as invoices, bills of lading and purchase orders placed in multipage files containing multiple documents without separators is described. The method is applicable in the presence of attachments and utilizes training data consisting of fields of interest and their locations in documents together with invariant fields frequently present in documents.