Character-Specific Learning Data Augmentation for OCR Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing techniques for character recognition, such as OCR, suffer from reduced recognition accuracy due to adverse effects of augmented learning data, particularly when background images like horizontal lines are added to specific characters, leading to erroneous recognition.
Innovation Solution
An image processing apparatus that generates learning data by applying different augmentation processing for specific characters, such as handwritten numbers or dates, to prevent synthesis with background images and restrict rotation angles, thereby improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If data augmentation is applied to improve recognition accuracy, then general recognition performance is improved, but specific characters (like '7') are erroneously recognized due to adverse effects from background image synthesis
Solution Approach 1:
The patent applies different augmentation strategies to different character types. Specifically, it identifies that certain characters (like '7', '1', '0') are adversely affected by background image synthesis, and applies restricted augmentation or no augmentation to these specific characters while maintaining full augmentation for other characters. This localized differentiation resolves the contradiction by preserving recognition accuracy improvement for most characters while protecting specific vulnerable characters from erroneous recognition.
Solution Approach 2:
The patent segments the character recognition task into different categories based on character susceptibility to augmentation artifacts. It separates characters into groups: those that benefit from background synthesis augmentation and those that are adversely affected. By segmenting the learning data generation process in this way, the system can apply appropriate augmentation strategies to each segment, thereby improving overall recognition accuracy while maintaining reliability for specific character types.
2Device complexity
If uniform augmentation processing is applied to all characters, then processing simplicity is maintained, but recognition accuracy for specific characters deteriorates
Solution Approach 1:
The patent implements local quality by applying different augmentation processing to different character types. It identifies specific characters (such as '7', '1', '0') that are adversely affected by certain augmentation operations and applies restricted or alternative augmentation to these characters, while maintaining standard augmentation for other characters. This approach increases processing complexity only where necessary to improve recognition accuracy for vulnerable characters.
Solution Approach 2:
The patent changes the augmentation parameters based on character type. For characters susceptible to erroneous recognition, it modifies the augmentation parameters (e.g., disables background image synthesis, restricts rotation angles) while maintaining standard parameters for other characters. This parameter differentiation resolves the contradiction by tailoring the processing to each character's specific recognition requirements.
Data Source
AI summary
An image processing apparatus includes an acquisition processing unit that acquires character image data, and a generation processing unit that generates learning data by executing predetermined augmentation processing on the character image data. In a case where the character image data is a specific character, the generation processing unit generates learning data by executing, on the specific character, augmentation processing different from augmentation processing for character image data other than the specific character.