OCR Training Data Generation via UI Font Rendering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional optical character recognition (OCR) systems face challenges in accurately determining bounding boxes for characters in images, especially when dealing with unknown fonts or localized text, leading to manual correction needs and suboptimal OCR model accuracy due to errors in bounding box inference.
Innovation Solution
A method that generates training data by displaying characters in a user interface with specified font type and size, capturing images to define accurate bounding box dimensions and positions, and stitching these images into a combined training image with associated 'golden' text, eliminating the need for manual correction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If conventional OCR systems use automated bounding box inference based on image features alone, then the process is fully automated, but the accuracy of bounding box determination deteriorates especially for unknown fonts or localized text
Solution Approach 1:
The patent introduces an intermediary component (bounding box verification module) that mediates between automated inference and final bounding box determination. This module uses multiple inference methods including template matching, contour analysis, and language model constraints to verify and correct automated bounding boxes, thereby maintaining automation while improving accuracy for unknown fonts and localized text
Solution Approach 2:
The system implements feedback loops where initially generated bounding boxes are evaluated against multiple criteria (image contours, language models, contextual information). Incorrect bounding boxes are detected and corrected through iterative refinement processes, allowing the system to learn from errors and improve subsequent bounding box determinations without manual intervention
2Measurement precision
If manual verification is used to ensure accurate bounding boxes and text representation, then accuracy improves, but the time and labor required increase significantly
Solution Approach 1:
The system enables self-service verification where the OCR system automatically validates its own output through multiple internal checks including character recognition confidence scores, bounding box consistency verification, and language model validation. This automated self-verification maintains high accuracy while eliminating the need for extensive manual checking
Solution Approach 2:
The patent performs preliminary verification actions during the OCR processing pipeline itself, validating bounding boxes and text recognition results before final output. By incorporating verification steps early in the process rather than requiring post-processing manual review, the system achieves high accuracy while minimizing time loss
3Quantity of substance
If thousands of characters are processed in a single OCR training image, then the dataset coverage improves, but the manual labor needed to verify each character increases
Solution Approach 1:
The patent segments the verification process into modular components that can independently validate different aspects of character recognition (bounding box accuracy, character identification, text formatting). This segmentation allows parallel processing of verification tasks across thousands of characters, maintaining dataset coverage while dramatically improving verification throughput through automated specialized validators
Data Source
AI summary
Systems of the present disclosure generate accurate training data for optical character recognition (OCR). Systems disclosed herein generates images of a text passage as displayed piecemeal in a user interface (UI) element rendered in a selected font type and size, determine accurate dimensions and locations of bounding boxes for each character pictured in the images, stitch together a training image by concatenating the images, and associate the training image, the bounding box dimensions and locations, and the text passage together in a collection of training data. The collection of training data also includes a computer-readable master copy of the text passage with newline characters inserted therein.