Optical Word Conversion for Script-Based Languages
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document image-to-text technologies face challenges in accurately converting text from script-based languages like Arabic, Persian, Pashto, Urdu, and Thai due to their inability to utilize effective image comparison techniques, leading to inaccurate output.
Innovation Solution
The development of optical word conversion (OWC) techniques using machine learning, which involves whole word image comparison with word templates in a conversion database, filtering the search space, and constructing word confusion networks to produce accurate Unicode strings, while minimizing language-specific adaptations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image-to-text techniques are used for script-based languages, then the conversion process can be performed, but the output accuracy deteriorates due to inability to utilize accurate image comparison techniques
Solution Approach 1:
The patent segments the text conversion process into distinct stages: line segmentation to identify horizontal clusters, word segmentation to produce word segments from clusters, and character segmentation to identify individual characters. This segmentation allows each stage to be optimized independently, improving overall accuracy for script-based languages while maintaining reliable output through systematic processing
Solution Approach 2:
The patent introduces word templates as intermediary elements that serve as reference standards for comparison. These templates are created from ground truth data and stored in a database, acting as mediators between the input image and the conversion output. The templates enable accurate image comparison techniques to be applied, directly addressing the reliability issue in traditional approaches
2Measurement precision
If extensive whole word comparisons are performed without filtering, then conversion accuracy improves, but processing time increases significantly
Solution Approach 1:
The patent performs preliminary actions by creating word templates from ground truth data before the actual conversion process. These pre-prepared templates are stored in a database with their corresponding text representations. During conversion, the system retrieves and compares against these pre-existing templates rather than creating comparisons from scratch, significantly reducing processing time while maintaining accuracy
Solution Approach 2:
The patent applies different processing qualities to different parts of the conversion process. Word templates that are frequently used or more distinctive are given higher priority in the comparison process. The system can adjust the string similarity threshold locally based on the specific language and context, allowing accurate comparisons where needed while skipping less critical comparisons
Data Source
AI summary
The disclosed invention provides system and method for whole word conversion of text in an image. The system and method may convert the text in the image into Unicode strings. The system and method utilizes comparable features of word segments that are extracted from the text in the image. The comparable features of the word segments are created by mirroring comparable features of word templates stored in a conversion database. Word templates in the database are filtered by comparing the comparable features of the word segments with comparable features of the word templates in the database. The word segments are correlated to matching word candidates. Word confusion networks are constructed with the matching word candidates, and the word confusion networks are evaluated to output matching words for the word segments. This system and method reduce the search space in the database improving search efficiency.