OCR Training Image Generation for Partial Handwritten Character Entry
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to efficiently generate training data for optical character recognition (OCR) of handwritten characters that protrude from document frames, particularly due to the difficulty in collecting and augmenting images of characters entering from various directions, which affects recognition accuracy.
Innovation Solution
An image processing apparatus that adds noise to character images near their ends to simulate partial entry from other rows, generating training data by associating these modified images with correct answer classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation is performed by adding noises to random positions within a character image, then the quantity of training data is increased, but the effectiveness for generating part-entered images is insufficient
Solution Approach 1:
The patent applies local quality by changing the noise addition strategy from random positions to specific end portions of character images. By concentrating noise addition at the ends where characters most likely to protrude, the augmentation process generates more effective part-entered images that accurately represent real-world scenarios where characters extend beyond their designated entry fields.
Solution Approach 2:
The patent implements preliminary action by proactively adding noise to simulate character protrusion before the actual OCR recognition process. This pre-simulation of part-entered images allows the training model to learn from artificially generated but realistic protrusion cases, improving its ability to handle such situations during actual document processing.
2Ease of operation
If character row separation processing is performed to separate character blocks, then the structure of scanned images is organized, but character blocks may be entered by part of characters from other rows
Solution Approach 1:
The patent applies preliminary action by pre-training the OCR model with part-entered images that simulate characters protruding from other rows. This preparation allows the model to anticipate and correctly handle character blocks that contain partial intrusions, maintaining high recognition accuracy even when character row separation processing results in blocks entered by characters from adjacent rows.
3Measurement precision
If training is performed with comprehensive images including various entering aspects, then the recognition accuracy for part-entered images is improved, but the difficulty of collecting and creating such images increases
Solution Approach 1:
The patent applies copying by creating artificial copies of character images with added noise to simulate part-entered scenarios. Instead of manually collecting diverse real-world examples of character protrusion, the system generates synthetic training images by copying existing character images and adding noise at end portions, efficiently creating comprehensive training data covering various entering aspects.
Solution Approach 2:
The patent implements parameter changes by modifying the noise addition parameters to specifically target end portions of character images. By adjusting the noise addition strategy to focus on terminal regions where protrusion occurs, the system efficiently generates training images that capture various entering aspects without requiring complex manual image collection processes.
Data Source
AI summary
A training image that simulates a character block entered by part of a character in another row is generated efficiently. A noise is added in the vicinity of an end portion of a character image so that a noise that seems to be caused by entering of part of a character in another row is reproduced for the character image representing a handwritten character.