Generic-Form OCR Domain Adaptation for Key-Value Pairing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing optical text recognition (OCR) systems struggle to accurately determine which text corresponds to which fields in forms without requiring specific location definitions and uniform form formats, necessitating time-consuming and costly manual effort for ground-truth dataset generation.
Innovation Solution
A method for optical text recognition on generic forms involves generating line grouping and feature extraction rules, using a key-value classifier model to determine the probability of lines being keys or values, and a key-value pairing model to predict corresponding values, based on training from generic forms without specific ground truth.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If OCR is performed on forms without specific location definitions and uniform formats, then the system's adaptability to different form types is improved, but the measurement precision of text-to-field correspondence deteriorates
Solution Approach 1:
The patent transforms the form recognition problem from location-based to feature-based by changing the parameters used for identification. Instead of requiring fixed locations and uniform formats, the system extracts features such as line groupings, key-value patterns, and structural relationships that vary across different form types, enabling accurate text-to-field correspondence without location constraints
Solution Approach 2:
The patent introduces feature extraction and line grouping as intermediary steps between OCR text extraction and field identification. These intermediaries transform raw OCR output into structured representations that capture the logical relationships between text and fields, bridging the gap between adaptable form processing and precise correspondence determination
2Measurement precision
If manual ground-truth dataset generation is performed to improve OCR accuracy, then the measurement precision is improved, but the loss of time and productivity deteriorates
Solution Approach 1:
The patent implements self-service by enabling the system to automatically generate its own training data and ground truth annotations through unsupervised feature extraction and pattern recognition. The system processes generic forms to learn line grouping rules, key-value relationships, and field structures without requiring manual annotation, thereby achieving high OCR accuracy while eliminating time-consuming manual dataset generation
Solution Approach 2:
The patent performs preliminary feature extraction, line grouping, and structural analysis on generic forms before actual OCR processing. By pre-learning form structures, key-value patterns, and field relationships from unlabeled data, the system prepares accurate models in advance that enable precise text-to-field correspondence without requiring manual ground truth during deployment
3Measurement precision
If manual ground-truth dataset generation is performed to improve OCR accuracy, then the measurement precision is improved, but the cost of dataset creation deteriorates
Solution Approach 1:
The system automatically generates training datasets and ground truth annotations by processing generic forms through unsupervised feature extraction and pattern recognition, eliminating the need for expensive manual annotation services while maintaining high OCR accuracy
Solution Approach 2:
The patent uses generic forms as templates that can be copied and adapted to various specific form types. By learning structural patterns from a small set of generic forms, the system generalizes to numerous specific forms without requiring separate manual annotation for each, thereby reducing the overall cost of dataset creation
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The disclosed technology is generally directed to optical text recognition for forms. In one example of the technology, line grouping rules are generated based on the generic forms and a ground truth for the generic forms. Line groupings are applied to the generic forms based on the line grouping rules. Feature extraction rules are generated. Features are extracted from the generic forms based on the feature extraction rules. A key-value classifier model is generated, such that the key-value classifier model is configured to determine, for each line of a form: a probability that the line is a value, and a probability that the line is a key. A key-value pairing model is generated, such that the key-value pairing model is configured to predict, for each key in a form, which value in the form corresponds to the key.