Table Structure Recognition in Distorted Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods fail to accurately analyze the structure of tables in images with geometric distortion, such as those captured by cameras, as they rely on grid structures and spacing between words, which are not applicable to distorted images.
Innovation Solution
A method and apparatus that extract lines forming the table, determine line intersections using crossing functions, identify crossing models based on shapes, and generate data describing the table structure, including cells, to recognize and generate table structures in geometrically distorted images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional methods relying on grid structures and word spacing are used, then table structure analysis is simple and fast, but accuracy deteriorates when geometric distortion is present
Solution Approach 1:
The patent changes the fundamental parameters for table detection from fixed grid coordinates and uniform spacing to geometric transformation-invariant line extraction and intersection detection. By using perspective transformation models and adaptive grid generation based on detected table corners, the system maintains accuracy under geometric distortion while managing complexity through structured parameter adaptation.
Solution Approach 2:
The patent introduces an intermediary transformation layer that maps distorted image coordinates to a normalized reference frame. By detecting table corners and applying perspective transformation to warp the distorted table structure into a standard grid, the system enables accurate cell and content recognition while separating the complex distortion handling from the simple grid-based processing.
2Measurement precision
If line extraction and intersection detection are performed, then accuracy in distorted images is improved, but processing time increases
Solution Approach 1:
The patent performs preliminary actions by pre-defining the set of possible crossing models and their geometric characteristics before processing. By preparing lookup tables for crossing patterns and pre-computing geometric constraints, the system can quickly match detected line intersections against predefined models rather than performing complex real-time analysis, thus reducing processing time while maintaining accuracy.
Solution Approach 2:
The patent segments the table structure recognition into distinct independent stages: corner point detection, line extraction, intersection identification, crossing model matching, and cell generation. Each stage processes specific information independently and feeds results to the next stage, enabling efficient pipelined processing and reducing overall time by avoiding redundant computations across stages.
3Measurement precision
If multiple crossing models are considered for each intersection, then recognition accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent applies local quality by selecting and evaluating crossing models selectively based on local geometric characteristics at each intersection. Instead of uniformly considering all possible crossing models everywhere, the system analyzes the specific geometric context (line orientations, angles, and spatial relationships) at each intersection to determine which crossing models are locally appropriate, reducing computation while maintaining accuracy.
Solution Approach 2:
The patent uses partial action by considering only a subset of crossing models that are geometrically plausible at each intersection rather than exhaustively evaluating all possible models. By filtering crossing models based on pre-computed geometric constraints and visual evidence, the system performs sufficient analysis to achieve accurate recognition without the excessive computation of checking every possible crossing configuration.
Data Source
AI summary
A method of recognizing and generating a structure of a table included in an image is provided. The method includes extracting lines forming the table from among connection components forming an image, determining line intersections by using crossing functions matched with the lines, determining one of a plurality of crossing models identified based on a plurality of crossing shapes, in correspondence with each of the line intersections, and generating data about the table, which includes at least one cell determined using the determined crossing model.