Selective Image Line Removal for Higher OCR Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing optical character recognition (OCR) techniques struggle with line removal from images, as removing lines that overlap with text often distorts the text and reduces OCR accuracy, necessitating additional processing or manual review, which consumes computing resources.
Innovation Solution
An image processing platform identifies first lines in a first dimension and second lines in an orthogonal dimension, determining non-intersecting portions of the first lines to selectively remove them, thereby minimizing text disturbance and enhancing OCR accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If lines are removed from the image to improve OCR accuracy, then OCR accuracy is improved, but text distortion increases and computing resources are consumed
Solution Approach 1:
The patent segments line removal into two independent passes: first identifying lines in a horizontal direction using a first bounding box, then identifying lines in a vertical direction using a second bounding box. This segmentation allows selective removal of only the harmful horizontal lines while preserving vertical text structures, thereby improving OCR accuracy without causing text distortion.
Solution Approach 2:
The patent applies different processing qualities to different regions: it removes horizontal lines that extend beyond text boundaries (identified through bounding box analysis) while preserving vertical lines and text content. This local differentiation ensures that only the harmful elements are removed, maintaining text integrity and avoiding distortion.
2Measurement precision
If additional processing is performed to handle line removal, then OCR accuracy improves, but computing resources are consumed
Solution Approach 1:
The patent performs preliminary line identification and classification before the actual removal process. By first analyzing the image with two bounding boxes to identify which lines are horizontal versus vertical, and which horizontal lines extend beyond text boundaries, the system prepares a selective removal list. This preliminary action enables efficient processing by avoiding unnecessary computations on regions that don't require modification.
Solution Approach 2:
The patent applies partial action by removing only the specific subset of horizontal lines that extend beyond text boundaries (identified through bounding box analysis), rather than removing all lines or requiring complete manual review. This partial removal approach achieves sufficient OCR accuracy improvement without the excessive computing resources that would be required for complete processing or manual verification.
3Measurement precision
If manual review is required for line removal, then OCR accuracy can be maintained, but productivity decreases
Solution Approach 1:
The patent implements self-service by automatically identifying and removing harmful horizontal lines through a systematic process using two bounding boxes. The system autonomously determines which lines to remove based on their spatial relationships with text boundaries, eliminating the need for manual review while maintaining high OCR accuracy. This automation directly improves processing throughput.
Solution Approach 2:
The patent incorporates feedback through iterative bounding box analysis: first bounding boxes identify horizontal lines, second bounding boxes verify vertical line positions, and the system uses this feedback to selectively remove only the harmful horizontal lines. This feedback mechanism ensures accuracy while maintaining automation, preventing the need for manual review and preserving high productivity.
Data Source
AI summary
In some implementations, a device may process an image to identify one or more first lines of the image that extend in a first dimension. The device may process the image to identify one or more second lines of the image that extend in a second dimension orthogonal to the first dimension. The device may identify portions of the one or more first lines that do not intersect with the one or more second lines. The device may process the image to obtain a version of the image in which the portions of the one or more first lines are removed.