OCR Table Detection for Rotated and Skewed Documents
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OCR systems face challenges in efficiently identifying and extracting table structures from documents, particularly when tables are rotated or skewed, due to the resource-intensive nature of large object detection models and limitations in training data.
Innovation Solution
A transformer-based table detection model that operates on OCR data, utilizing an attention mechanism to identify correlations between words and generate bounding boxes for table elements, reducing the need for additional processing resources and improving accuracy on skewed or rotated tables.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large object detection models are used to identify tables in document images, then table detection capability is improved, but processing resources and time consumption increase significantly
Solution Approach 1:
The patent segments the table detection task into multiple stages: first using a lightweight model to detect potential table regions, then applying more sophisticated analysis only to those specific regions. This segmentation allows the system to achieve accurate table detection while avoiding the computational burden of applying large detection models to entire document images, thus resolving the contradiction between detection capability and processing time.
Solution Approach 2:
The patent performs preliminary processing by detecting text regions and structures before final table identification. By pre-processing the document to identify potential table areas using simpler, faster methods, the system reduces the computational load on subsequent table detection steps, thereby maintaining accuracy while reducing processing time.
2Measurement precision
If large object detection models are deployed for table identification, then detection accuracy is improved, but processing resources and hardware requirements increase
Solution Approach 1:
The patent divides the detection system into hierarchical levels, using lightweight models for initial screening and more sophisticated models only where needed. This segmentation enables accurate table detection to be achieved without deploying complex large models across the entire system, thus reducing hardware requirements while maintaining detection accuracy.
Solution Approach 2:
The patent introduces intermediate processing steps and auxiliary models that act as mediators between the input document and the final table detection output. These intermediaries pre-process the data and filter potential table regions, allowing the main detection model to operate more efficiently with reduced computational demands on hardware.
3Productivity
If traditional OCR methods are used for table detection, then resource consumption is reduced, but accuracy on rotated or skewed tables deteriorates
Solution Approach 1:
The patent implements dynamic adaptability in the detection system by incorporating transformations and adjustments that allow the model to handle various table orientations. The system dynamically adjusts its detection parameters and coordinate transformations based on the detected table characteristics, enabling accurate detection of rotated or skewed tables while maintaining resource efficiency through the use of lightweight models.
Solution Approach 2:
The patent changes key parameters such as coordinate systems, rotation angles, and scaling factors to accommodate rotated or skewed tables. By dynamically adjusting these parameters based on the input document characteristics, the system maintains high accuracy for non-standard table orientations while using computationally efficient methods.
Data Source
AI summary
Various examples are directed to systems and methods for determining table data from a document image depicting a plurality of words and at least one table comprising at least a portion of the plurality of words. For example, Optical Character Recognition (OCR) data may be determined based on the document image. A table detection model may be executed based at least in part on the OCR data.