Table Recognition Feature Fusion for Blank Cells and OCR Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional table recognition methods suffer from low accuracy due to the lack of information exchange between table structure recognition and character content recognition, leading to difficulties in identifying blank regions in table cells and distinguishing characters like '10' and 'IO'.
Innovation Solution
A multi-modal information-based table recognition method that fuses image vision features and character content features to improve accuracy by utilizing the information from both tasks, including neural networks and optical character recognition (OCR) to enhance cell detection and character recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional separate task recognition is used (first table structure, then character content), then the process is simple and straightforward, but the recognition accuracy is low
Solution Approach 1:
The patent merges table structure recognition and character content recognition into a unified multi-task learning framework. The neural network simultaneously performs both recognition tasks by sharing common feature extraction layers, allowing information to be exchanged between tasks. This integration enables the system to leverage contextual information from both structures and characters together, thereby improving overall recognition accuracy while maintaining a unified model architecture.
Solution Approach 2:
The patent implements a multi-task learning model that serves multiple functions: it recognizes both table structure (rows, columns, cells) and character content within the same neural network framework. The shared feature extraction layers process image data to generate representations that are useful for both tasks, making the system universally capable of handling both recognition objectives simultaneously rather than requiring separate specialized systems.
2Loss of information
If separate task recognition is used, then the implementation is straightforward, but information exchange between tasks is lost
Solution Approach 1:
The patent combines table structure recognition and character content recognition into a single multi-task learning model. The shared neural network layers process the input image and generate feature representations that contain information useful for both tasks. This merging allows bidirectional information exchange: structure information helps characterize content regions, while character information refines structure boundaries, thereby reducing information loss that would occur in separate processing pipelines.
3Measurement precision
If conventional methods are used, then processing is faster and simpler, but accuracy in identifying blank regions and distinguishing similar characters is poor
Solution Approach 1:
The patent employs preliminary feature extraction that simultaneously prepares information for both table structure recognition and character content recognition. The shared neural network layers pre-process the image to generate comprehensive feature representations before the final task-specific predictions. This preliminary multi-purpose feature extraction enables more accurate cell detection and character distinction by considering both structural and textual information together from the outset, rather than sequentially.
Data Source
AI summary
Disclosed are a table recognition method and apparatus. The table recognition method includes steps of obtaining an image vision feature and a character content feature of a table image; fusing the image vision feature and the character content feature of the table image to acquire a first fusion feature, and carrying out recognition based on the first fusion feature to acquire a table structure; and performing, based on the table structure, character recognition on the table image to acquire table character contents.