Multi-Engine OCR Error Correction via Probability Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current optical character recognition (OCR) engines produce errors in digitized text due to imperfections in documents and scanning processes, which are not adequately corrected by existing OCR applications, particularly for applications like Text-to-Speech and Republishing, where higher accuracy is required.
Innovation Solution
A method involving multiple OCR engines processing text images, with an error probability estimator comparing outputs, grouping mismatches into clusters, and presenting questions to human operators for correction, while actively training the estimator using operator answers to improve accuracy and reduce false positives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple OCR engines are used to process text images, then OCR output quality is improved, but device complexity and processing time increase
Solution Approach 1:
The patent combines multiple OCR engines to process the same text image, comparing their outputs to identify and correct errors. By merging the results from multiple engines and using an error probability estimator to evaluate discrepancies, the system achieves higher accuracy than any single engine could provide alone.
Solution Approach 2:
The error probability estimator acts as an intermediary component that receives outputs from multiple OCR engines, evaluates the reliability of each engine's output, and determines which results to trust. This mediator resolves the complexity by providing a systematic method to integrate multiple engine outputs without requiring manual intervention for every discrepancy.
2Measurement precision
If human operators review all OCR mismatches, then accuracy is improved, but productivity and cost increase due to extensive manual intervention
Solution Approach 1:
Instead of having human operators review all OCR outputs uniformly, the system applies local quality control by directing human attention only to specific areas where errors are most likely to occur. The error probability estimator identifies high-risk mismatches, and human operators review only those cases, rather than examining every single output.
Solution Approach 2:
The system performs partial manual review by having human operators examine only a subset of OCR mismatches—specifically those with high error probabilities identified by the estimator. This partial action approach achieves high accuracy without requiring exhaustive manual review of all outputs, thereby maintaining productivity.
3Measurement precision
If the error probability estimator is trained extensively with ground truths, then measurement precision is improved, but loss of time and computational resources increase
Solution Approach 1:
The error probability estimator is pre-trained using ground truth data from previously processed text images. This preliminary training establishes a baseline model that can quickly evaluate error probabilities without requiring extensive training each time new text images are processed. The system performs the time-consuming training work in advance.
Solution Approach 2:
The error probability estimator continuously improves itself by learning from human operator corrections. As human operators review and correct OCR mismatches, the system uses this feedback to refine its error probability estimates, progressively reducing the need for extensive manual review over time. The system serves itself by automatically learning from its operations.
Data Source
AI summary
OCR errors are identified and corrected through learning. An error probability estimator is trained using ground truths to learn error probability estimation. Multiple OCR engines process a text image, and convert it into texts. The error probability estimator compares the outcomes of the multiple OCR engines for mismatches, and determines an error probability for each of the mismatches. If the error probability of a mismatch exceeds an error probability threshold, a suspect is generated and grouped together with similar suspects in a cluster. A question for the cluster is generated and rendered to a human operator for answering. The answer from the human operator is then applied to all suspects in the cluster to correct OCR errors in the resulting text. The answer is also used to further train the error probability estimator.