Document Image Recognition via Side Profile Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document image processing techniques face challenges in accurately recognizing tables, flowcharts, and text, especially in hand-written documents with deformed shapes, due to instability and computational complexity, particularly when dealing with non-straight lines and overlapping components.
Innovation Solution
A method that calculates side profiles for document images, removes boundaries based on line widths, and re-calculates profiles to differentiate between tables, text, and flowcharts by analyzing smoothness, straight line fits, and profile differences, allowing for accurate classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If line detector based approach is used for table recognition, then straight lines can be detected, but the method becomes computationally heavy and unstable for hand-written documents with non-straight lines
Solution Approach 1:
The patent segments the document image into multiple profiles by scanning horizontal and vertical lines, creating separate one-dimensional representations that simplify the detection of tabular structures while reducing computational burden compared to full image analysis
Solution Approach 2:
The patent transforms the two-dimensional image problem into one-dimensional profile analysis by creating horizontal and vertical scans, enabling simpler detection algorithms to identify tabular structures through pattern recognition in the profile data
2Adaptability or versatility
If connected component based method is used, then empty tables can be handled, but the method becomes unstable when text overlaps table line segments
Solution Approach 1:
The patent segments the image into horizontal and vertical profiles that separate line structure detection from text recognition, allowing tables to be identified through profile patterns even when text overlaps with table boundaries
Solution Approach 2:
The patent introduces profiles as an intermediary representation between the original image and the final recognition result, enabling stable detection of tabular structures through profile characteristics that are independent of text content
3Ease of manufacture
If conventional bottom-up recognition approach is used, then basic components can be extracted, but the method requires many restrictions such as handling only convex polygons and relatively straight lines
Solution Approach 1:
The patent creates one-dimensional profiles from the two-dimensional image, enabling the detection of arbitrary shapes including non-convex and deformed structures through pattern recognition in the profile data without requiring geometric constraints
Solution Approach 2:
The patent changes the representation parameters from geometric primitives (lines, polygons) to profile intensity patterns, enabling recognition of deformed and hand-written structures that do not conform to strict geometric rules
Data Source
AI summary
A method for recognizing a binary document image as a table, pure text, or flowchart includes calculating a side profile of the image for each of the four sides, calculating a boundary removal size N corresponding to each side based on widths of lines or strokes closest to that side, and for each side, removing a boundary of size N from the document image, and re-calculating the side profile for each side after the removal. Then, based on a comparison of the side profiles and the re-calculated side profiles, the input document image is recognized as a table if all side profiles change from smooth to non-smooth, as pure text if the side profile changes are small, and as a flowchart if the original side profiles contain multiple sharp changes and wide flat regions and if the side profile changes significantly in the previously wide flat regions.