Neighbor Embedding for OCR Text Super-Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image super-resolution techniques, particularly for text images, suffer from high character error rates and computational inefficiency when used for OCR enhancement, as they require large training models and are limited to specific font styles and sizes, making them unsuitable for diverse text documents.
Innovation Solution
A computationally efficient method using randomized k-d trees for neighbor embedding, which generates high-resolution images from low-resolution text images, reducing character error rates and allowing a single model to handle various Latin font styles and sizes, while requiring less memory than existing methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If exact nearest neighbor search is used in Chang et al's method, then high resolution image estimation is achieved, but computational speed is too slow for OCR enhancement
Solution Approach 1:
The patent uses approximate nearest neighbor search instead of exact nearest neighbor search, sacrificing a small amount of accuracy for dramatically improved computational speed. This allows the system to process images fast enough for OCR enhancement while maintaining sufficient quality
Solution Approach 2:
The patent changes the search parameter from exact to approximate nearest neighbor search, fundamentally altering the computational approach to achieve practical processing speeds while maintaining acceptable accuracy for OCR applications
2Measurement precision
If large training models are used in Chang et al's method, then accurate HR image estimates are obtained, but memory requirements are too high
Solution Approach 1:
The patent extracts only the essential features needed for text image super-resolution, using a simplified feature vector based on pixel intensities rather than complex derivative features. This reduces memory requirements while maintaining effectiveness for OCR enhancement
Solution Approach 2:
The patent uses a simpler feature representation that requires less memory to store training models, effectively creating a more compact version that achieves the same goal with reduced resource requirements
3Measurement precision
If Chang et al's method is used, then high resolution images are generated, but the method requires knowledge of font style and size, limiting versatility
Solution Approach 1:
The patent creates a universal super-resolution method that works across different font styles and sizes without requiring specific knowledge of the input. The system is designed to be font-agnostic, making it applicable to diverse text documents for OCR enhancement
Solution Approach 2:
The patent changes the approach from font-specific parameter matching to a general-purpose super-resolution method that adapts to different fonts through learning, removing the constraint of requiring known font characteristics
4Productivity
If interpolation methods are used for super-resolution, then computational speed is fast, but character error rates in OCR are unacceptably high
Solution Approach 1:
The patent replaces simple interpolation mechanics with a learning-based neighbor embedding approach that learns from training data. This substitution maintains reasonable computational speed while dramatically improving OCR accuracy by capturing complex text patterns
Data Source
AI summary
Single-image super-resolution (SISR) is the problem of generating a high resolution image from a single low resolution image. The SISR technique known as neighbor embedding utilizes a training ensemble of pairs of low and high resolution image patches, where the patches in a given pair represent the same image region. The present invention improves upon prior neighbor embedding algorithms by offering a practical, computationally efficient method of neighbor embedding for generating a high resolution version of a low resolution image. The technique may also be applied to generate high resolution versions of low resolution text images for subsequent input into OCR engines. OCR character error rates found on the high resolution images are drastically lower than those found when OCR is applied to the original low resolution text images.