ML Text Segmentation for OCR Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Optical Character Recognition (OCR) from scanned documents faces challenges due to poor image quality, scan noise, and the presence of graphics and tables, which confuse OCR techniques, and existing pre-processing methods often fail to accurately differentiate text pixels from non-text pixels.
Innovation Solution
A machine learning (ML) text segmentation model is trained to classify each pixel as text or non-text, providing a per-pixel image classification map that is used as input to an OCR engine, rather than applying transformations to image pixels, thereby improving OCR accuracy by focusing on the skeletal structure of characters and providing metadata for better text recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional OCR is applied directly to scanned document images, then the OCR process can be performed, but accuracy deteriorates due to scan noise, poor image quality, and confusion from graphics and tables
Solution Approach 1:
The patent applies preliminary action by performing text segmentation and generating a per-pixel classification map before the OCR process. The ML model pre-processes the image to identify and segment text regions, creating a cleaned classification map that removes scan noise and distinguishes text from non-text elements (graphics, tables) before OCR is applied, thereby improving accuracy without requiring changes to the OCR engine itself
Solution Approach 2:
The patent introduces an intermediary component - the per-pixel classification map generated by the ML model - that acts as a mediator between the original scanned image and the OCR engine. This intermediary map filters out harmful factors (noise, graphics, tables) while preserving text information, allowing the OCR engine to process only the relevant text regions with improved quality
2Measurement precision
If pre-processing transformations are applied directly to image pixels, then some improvement in text visibility may be achieved, but accuracy deteriorates because existing methods fail to accurately differentiate text pixels from non-text pixels
Solution Approach 1:
The patent applies local quality by using an ML model to analyze each pixel individually and assign it a classification (text or non-text) based on local patterns and context. The per-pixel classification map assigns different properties to different regions of the image, allowing precise differentiation of text pixels from non-text pixels while preserving the unique characteristics of each region, thereby improving both accuracy and reliability
Solution Approach 2:
The patent changes the parameter representation of the image by transforming it from raw pixel values to a per-pixel classification map with discrete text/non-text labels. This parameter transformation allows the system to accurately differentiate text from non-text elements by changing the fundamental representation from continuous grayscale/color values to discrete semantic classifications
Data Source
AI summary
Embodiments regard improving Optical Character Recognition (OCR) performance. A method includes providing an image including text as input to a text segmentation model, receiving, from the text segmentation model, a per-pixel segmentation map of the image, providing the per-pixel segmentation map as input to an OCR engine, and receiving, as output from the OCR engine based on the per-pixel segmentation map, a digitized version of the image.