Dynamic OCR Character Resolution via Context-Aware Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current OCR systems lack perfect accuracy and rely on static voting protocols to combine outputs from multiple engines, which do not adapt to changing performance in specific document contexts.
Innovation Solution
A dynamic OCR system that models the performance of multiple OCR engines over time, using a character resolution model to resolve differences in outputs based on document context and updates this model with new information to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple OCR engines are used with static voting protocols, then character recognition coverage is improved, but adaptability to different document contexts deteriorates
Solution Approach 1:
The patent implements dynamic voting protocols that adapt to different document contexts by analyzing document characteristics (such as document type, quality metrics, and content features) and adjusting the voting weights and selection criteria accordingly. This transforms the static voting mechanism into a dynamic system that can optimize its behavior based on the specific context, thereby resolving the contradiction between maintaining reliable multi-engine recognition and adapting to diverse document types.
Solution Approach 2:
The system changes key parameters of the voting protocol (such as weight assignments to different engines, threshold values, and selection criteria) based on document context analysis. By modifying these parameters dynamically according to document characteristics, the system achieves both high reliability through multi-engine validation and adaptability to different document contexts.
2Device complexity
If static voting protocols are used to combine OCR outputs, then system complexity is reduced, but recognition accuracy in specific contexts deteriorates
Solution Approach 1:
The patent applies local quality by implementing context-specific voting strategies tailored to different document types and quality levels. Instead of a uniform voting protocol, the system analyzes local document characteristics (such as document category, image quality metrics, and content features) and applies appropriate voting parameters locally, thereby improving accuracy without requiring excessively complex global protocols.
Solution Approach 2:
The system performs preliminary analysis of document characteristics before applying the voting protocol. By pre-processing the document to identify its type, quality level, and other relevant features, the system can select and configure the appropriate voting strategy in advance, improving recognition accuracy while keeping the actual voting mechanism relatively simple.
3Reliability
If multiple OCR engines operate in parallel, then character recognition coverage is improved, but processing time increases
Solution Approach 1:
The patent implements a staged voting approach where OCR engines are executed in a controlled sequence or with selective parallelism based on document characteristics. For high-confidence cases or simple documents, fewer engines may be activated (partial action), while more complex documents trigger full multi-engine processing. This balances recognition coverage with processing time efficiency.
Solution Approach 2:
The system employs periodic execution of additional OCR engines based on confidence thresholds and document complexity assessments. Instead of always running all engines in parallel, the system periodically evaluates whether additional engine execution is necessary based on intermediate results and document features, thereby reducing unnecessary processing time while maintaining adequate recognition coverage.
Data Source
AI summary
Optical character recognition is described in various implementations. In one example implementation, a method may include receiving a plurality of optical character recognition (OCR) outputs provided by a respective plurality of OCR engines, each of the plurality of OCR outputs being representative of text depicted in a portion of an electronic image. The method may also include identifying a document context associated with the electronic image, and generating an output character set by applying a character resolution model to resolve differences among the plurality of OCR outputs. The character resolution model may define a probability of character recognition accuracy for each of the plurality of OCR engines given the identified document context. The method may also include updating the character resolution model to generate an updated character resolution model such that subsequent generating of output character sets are based on the updated character resolution model.