Form Identification via Topological-Semantic Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In the field of document and form analysis, challenges arise from unstructured forms, scanning extraction errors, tables with varying sizes and locations, and the need for scalable and robust solutions that can handle rotation, translation, and scaling while maintaining accurate form recognition and registration.
Innovation Solution
A machine learning system is trained to identify generic regions on forms using topographical and semantic relationships, allowing it to recognize and categorize fields and sub-regions across different forms, even when they are rotated, translated, or scaled, by employing a deep learning module that updates weights based on incorrect categorizations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If semi-structured form representation is used to mix topological features and semantic information, then form recognition capability is improved, but understanding associations among features becomes challenging when forms undergo rotation, translation, and scaling
Solution Approach 1:
The patent segments the form processing task into distinct stages: first identifying topological features (lines, shapes, spatial relationships) independently of semantic content, then separately processing semantic information. This segmentation allows the system to handle geometric transformations (rotation, translation, scaling) at the topological level while maintaining semantic associations, resolving the contradiction between adaptability and complexity.
Solution Approach 2:
The patent introduces an intermediary representation layer that decouples topological features from semantic information. This intermediary structure uses geometric invariants and relational graphs to represent form structure independently of content, allowing the system to maintain feature associations even when forms undergo transformations, thus reducing the complexity of understanding associations while preserving recognition capability.
2Productivity
If traditional OCR or ICR scanning extraction is used, then text extraction is achieved, but scanning extraction errors occur reducing reliability
Solution Approach 1:
The patent applies preliminary action by performing topological feature extraction and form structure analysis before conducting OCR/ICR text recognition. By first establishing the geometric framework and spatial relationships of form elements, the system creates a robust structure that guides subsequent text extraction, reducing errors from misalignment or poor image quality while maintaining extraction efficiency.
Solution Approach 2:
The patent implements feedback mechanisms where the topological structure and semantic context obtained from preliminary analysis are fed back to guide and correct the OCR/ICR extraction process. This feedback loop allows the system to adjust extraction parameters based on form structure and use contextual information to resolve ambiguous character recognition, thereby improving reliability without sacrificing productivity.
3Adaptability or versatility
If tables with variable sizes and locations are processed, then form variability is accommodated, but consistent identification and alignment across different forms becomes difficult
Solution Approach 1:
The patent applies parameter changes by using geometric invariants and normalized coordinate systems that remain consistent regardless of table size, position, or orientation. Instead of relying on fixed absolute coordinates, the system transforms table parameters into relative, scale-invariant representations that maintain precise identification and alignment across forms with variable dimensions and layouts, resolving the contradiction between adaptability and measurement precision.
Data Source
AI summary
Aspects of the present invention provide a computer-implemented method of training a machine learning system to identify forms. In embodiment, the method may include: receiving a form as an input image; identifying one or more fields in the input image; for each identified field, identifying one or more sub-regions in the identified field; responsive to identification of the one or more fields, categorizing the one or more fields; identification of relative locations of the one or more fields in the input image; and, responsive to the identification of the relative locations, categorizing the form. Other aspects of the present invention provide a computer-implemented method of using a machine learning system to identify forms, using the just-enumerated method.