Lineless Table Data Extraction via Delaunay Triangulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods fail to accurately detect and extract structured information from lineless tables in scanned or raster images, as they cannot determine associations among characters or between characters and locations within a table, limiting their ability to process images without explicit structural information.
Innovation Solution
The method involves using optical character recognition (OCR) to detect characters and define bounding boxes, generating a graph through Delaunay triangulation, refining the graph to remove excess edges, and using a neural network to predict row and column labels, thereby identifying structural information in lineless tables without relying on explicit table lines.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional OCR methods are used to detect characters in scanned images, then character recognition is achieved, but the ability to determine associations among characters and identify table structure is lost
Solution Approach 1:
The patent segments the table detection problem into multiple stages: first performing OCR to detect individual characters, then constructing a graph where characters are nodes and potential table edges are connections, and finally refining this graph to identify the actual table structure. This segmentation allows the system to process character recognition and structure detection separately while maintaining their relationships.
Solution Approach 2:
The patent introduces a graph structure as an intermediary between OCR output and final table detection. The graph construction algorithm creates potential edges between characters based on spatial relationships, and the refinement process eliminates false edges. This intermediary structure preserves spatial information that would otherwise be lost in traditional OCR pipelines.
2Reliability
If a complete graph is constructed from all character bounding boxes, then all potential relationships are captured, but the graph contains excessive redundant edges that complicate processing
Solution Approach 1:
The patent extracts only the necessary edges from the complete graph through a refinement process. By analyzing spatial relationships and removing edges that do not correspond to actual table boundaries, the system retains only the meaningful connections between characters, eliminating redundant information while preserving structural accuracy.
Solution Approach 2:
The patent initially constructs a complete graph with all possible edges (excessive action) to ensure no potential relationships are missed, then applies refinement to remove redundant edges. This approach ensures completeness first, then optimizes for efficiency, allowing the system to capture all potential table structures before filtering out false positives.
3Measurement precision
If Delaunay triangulation is used to generate graph edges from character bounding boxes, then spatial relationships are captured, but the graph includes edges that do not represent actual table boundaries
Solution Approach 1:
The patent converts the harmful effect of false edges generated by Delaunay triangulation into a benefit by using these edges as candidates for refinement. The excessive edges provide comprehensive coverage of potential relationships, and the subsequent refinement process selectively retains only the accurate table boundary edges, transforming the initial over-generation into a thorough search space.
Data Source
AI summary
A method for extracting data from lineless tables includes storing an image including a table in a memory. A processor operably coupled to the memory identifies a plurality of text-based characters in the image, and defines multiple bounding boxes based on the characters. Each of the bounding boxes is uniquely associated with at least one of the text-based characters. A graph including multiple nodes and multiple edges is generated based on the bounding boxes, using a graph construction algorithm. At least one of the edges is identified for removal from the graph, and removed from the graph to produce a reduced graph. The reduced graph can be sent to a neural network to predict row labels and column labels for the table.