Table Structure Recognition via Deep Spatial Word Association
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Table Structure Recognition (TSR) techniques face challenges in handling complex tables with nesting, missing values, and overlapping columns due to their low generalization ability and high computational resource requirements, especially when dealing with scanned images lacking metadata.
Innovation Solution
A method and system for table structure recognition via deep spatial association of words, which identifies word pairs using immediate top and left neighbors, eliminates redundant pairs, and uses a single classifier to determine word associations, reducing computational complexity while maintaining accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all possible word pairs are generated to improve association insights, then the accuracy of table structure recognition is improved, but the computational complexity and time effort increase significantly
Solution Approach 1:
The patent extracts only the essential word pairs needed for table structure recognition by focusing on spatial relationships (horizontal and vertical neighbors) rather than generating all possible word pairs. This selective extraction maintains association accuracy while dramatically reducing computational complexity.
Solution Approach 2:
The patent applies local quality by considering only local spatial relationships between words (immediate neighbors in horizontal and vertical directions) rather than global word pair associations. This localized approach captures sufficient structural information while reducing the number of word pairs to a manageable level.
2Productivity
If a sampling approach is used to reduce the number of word pairs, then the computational resources are reduced, but the random selection produces skewed distribution leading to biased network and hampered performance
Solution Approach 1:
The patent changes the selection parameter from random sampling to spatially-aware sampling based on bounding box coordinates. By using spatial parameters (horizontal and vertical positions) to guide word pair selection, the method achieves both computational efficiency and model performance, avoiding the skewed distribution problem of random sampling.
3Measurement precision
If multiple classification models are used to determine word associations, then the accuracy of spatial association is improved, but the computational resources and complexity increase
Solution Approach 1:
The patent merges the functionality of multiple classification models into a single neural network model that processes both horizontal and vertical word pairs simultaneously. This unified model reduces computational resources while maintaining the accuracy benefits of considering multiple spatial relationships.
4Productivity
If fixed number of words are extracted using Monte-Carlo Sampling, then the number of word pairs is reduced, but the neighborhood property is not focused leading to affected accuracy of association
Solution Approach 1:
The patent changes the sampling parameter from fixed random selection to spatially-proximity-based selection. By using bounding box coordinate parameters to identify horizontal and vertical neighbors, the method ensures that extracted word pairs maintain meaningful spatial relationships, preserving association accuracy while achieving processing efficiency.
Data Source
AI summary
State of art techniques that utilize spatial association based Table structure Recognition (TSR) have limitation in selecting minimal but most informative word pairs to generate digital table representation. Embodiments herein provide a method and system for TSR from an table image via deep spatial association of words using optimal number of word pairs, analyzed by a single classifier to determine word association. The optimal number of word pairs are identified by utilizing immediate left neighbors and immediate top neighbors approach followed redundant word pair elimination, thus enabling accurate capture of structural feature of even complex table images via minimal word pairs. The reduced number of word pairs in combination with the single classifier trained to determine the word associations into classes comprising as same cell, same row, same column and unrelated, provides TSR pipeline with reduced computational complexity, consuming less resources still generating more accurate digital representation of complex tables.