OCR Correction with LSTM and Modified Edit Distance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Optical Character Recognition (OCR) techniques achieve high accuracy but often result in errors, particularly in eKYC processes, where names and addresses are frequently misinterpreted, leading to incorrect corrections when using direct search methods.
Innovation Solution
A system utilizing a trained Long short-term memory (LSTM) neural network language model and a modified edit distance process to evaluate OCR results, determining the need for correction and identifying the most similar text from a constructed name and address corpus, while enriching the corpus with corrected results to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If direct searching of name/address corpus is used to correct OCR results, then correction speed is improved, but correction accuracy deteriorates due to many correct results being wrongly corrected
Solution Approach 1:
The patent introduces an intermediary verification mechanism between the OCR result and the corpus search. The system first identifies potential corrections through modified edit distance, then verifies them using multiple criteria including language model probability, character similarity metrics, and frequency analysis before applying corrections. This intermediary verification layer prevents wrong corrections while maintaining efficient processing.
Solution Approach 2:
The patent modifies the traditional edit distance parameter by introducing a weighted modified edit distance that incorporates visual similarity weights for different character types. Additionally, the system dynamically adjusts correction thresholds based on confidence scores from language models and corpus frequency statistics, allowing adaptive parameter changes to balance speed and accuracy.
2Measurement precision
If corpus coverage is increased to improve correction accuracy, then more OCR errors can be corrected, but system complexity and processing time increase
Solution Approach 1:
The patent performs preliminary actions by pre-processing the corpus to extract and store key features such as character frequency distributions, n-gram patterns, and visual similarity matrices before the actual correction process. Language models are pre-trained on the corpus, and edit distance thresholds are pre-calculated for common character pairs. This preliminary preparation reduces the complexity and processing time during actual OCR correction operations.
3Measurement precision
If modified edit distance with visual similarity weights is used, then correction accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent applies local quality by assigning different visual similarity weights to different character positions and types. Instead of using uniform weights for all character comparisons, the system identifies locally important character positions (such as distinguishing characters in names) and applies higher weights to their similarity calculations. This localized approach improves accuracy for critical fields while reducing overall computational complexity by not uniformly applying complex calculations to all positions.
Data Source
AI summary
Disclosed herein are computer-implemented methods, computer-implemented systems, and non-transitory, computer-readable media for automatic Optical Character Recognition (OCR) correction. One computer-implemented method includes evaluating an OCR result using a trained Long short-term memory (LSTM) neural network language model to determine whether correction to the OCR result is required. If correction to the OCR result is required, a most similar text relative to the OCR result is determined from a name and address corpus using a modified edit distance technique. The OCR result is corrected with the determined most similar text.