Table Cell Region Detection via Multi-Resolution Image Rescaling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OCR systems fail to accurately detect and maintain the relational structure of tables in scanned documents, leading to incorrect information representation, as they struggle to identify cell regions and borders within tables.
Innovation Solution
A method and system that rescale the primary image document to generate secondary images, detect predefined table features such as L-shaped edges and T-junctions, and generate clusters of connected components to identify cell regions by modifying pixel values, ensuring accurate detection and preservation of table structure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If OCR systems are used to extract text from scanned documents, then text extraction capability is improved, but table structure detection capability deteriorates
Solution Approach 1:
The patent segments the table detection problem into multiple stages: first detecting candidate regions using L-shaped edge patterns, then identifying T-junctions and intersections, and finally assembling these components into complete table structures. This segmentation allows the system to specialize in different aspects of table detection rather than attempting to solve the entire problem in one step, thereby improving both text extraction accuracy and table structure detection reliability.
Solution Approach 2:
The patent introduces intermediate structures such as L-shaped edges, T-junctions, and intersections as mediators between the raw image data and the final table structure. These intermediate features serve as building blocks that bridge the gap between simple text detection and complex table structure recognition, enabling the system to reliably detect tables while maintaining text extraction capability.
2Measurement precision
If existing techniques extract text from tables, then text extraction is improved, but relational structure preservation deteriorates
Solution Approach 1:
The patent performs preliminary detection of table boundaries, rows, and columns before extracting the text content. By first identifying the structural framework of the table using L-shaped edges and junction patterns, the system preserves the relational structure information and then extracts text within these predefined boundaries, preventing loss of structural context.
Solution Approach 2:
The patent changes the detection parameters and features used at different stages of processing. It uses L-shaped edge detection for initial candidate region identification, then applies T-junction and intersection detection for structure verification, and finally uses these structural parameters to guide text extraction. This parameter transformation approach ensures both text accuracy and structure preservation.
3Adaptability or versatility
If table detection is performed on scanned documents with varying arrangements, then adaptability is improved, but detection precision deteriorates
Solution Approach 1:
The patent employs universal detection patterns such as L-shaped edges, T-junctions, and intersections that can identify table structures regardless of their specific arrangement or orientation. These multi-functional features work across different table layouts, making the system adaptable while maintaining precision through the consistent geometric properties of the detected patterns.
Data Source
AI summary
The present disclosure is related to field of machine learning and image processing, disclosing method and device for identifying cell region of table including cell borders from an image document. Table detecting system rescales a primary image document into plurality of secondary image documents of different size and resolution to detect plurality of candidate regions comprising predefined table features in each secondary image document. Further, for each candidate region, set of connected components are determined and the connected components corresponding to the IDs that are present in more than one set of the connected components are clustered. Subsequently, areas corresponding to the clusters that are determined to form a table are cropped from the primary image document and each cell region of the table is identified by modifying pixel values of the clusters of the connected components in the cropped area.