Image Recognition Model Training with Virtual Data Generation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing OCR methods face inefficiencies in training deep learning models due to the difficulty in acquiring large amounts of private data, particularly for financial, card, and bill types, which hampers training efficiency and recognition accuracy.

Innovation Solution

The method involves generating a large amount of virtual data based on a small amount of real data and using both types in the training process to improve the image recognition model's efficiency and accuracy, utilizing a deep learning network to extract text data from training images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a large amount of real training data is used to train the deep learning network, then the recognition accuracy is improved, but the difficulty of acquiring data increases due to privacy constraints

Engineering Contradiction:
Improverecognition accuracyVSAvoiddifficulty of acquiring data
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

The patent applies the copying principle by generating virtual training data that replicates the characteristics and formats of real private data without using the actual sensitive information. Virtual images are created that mimic the structure, layout, and visual properties of real financial, card, and bill images, providing sufficient training material while preserving data privacy and avoiding acquisition difficulties.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent extracts the essential structural and visual characteristics from real data to create virtual training samples. By separating the useful training features (layout, format, visual patterns) from the sensitive private content, the system can generate abundant training data that maintains recognition accuracy while eliminating privacy concerns and acquisition barriers.

Inventive Principle:
Principle #2Taking out (Extraction)

2Quantity of substance

If virtual data is generated to supplement training data, then the availability of training data is improved, but the training model complexity increases

Engineering Contradiction:
Improveavailability of training dataVSAvoidtraining model complexity
Core Design Contradiction:
Quantity of substanceVSDevice complexity

Solution Approach 1:

The patent applies parameter changes by systematically varying key parameters in the virtual data generation process, including image resolution, field positions, text content, and layout configurations. This approach generates diverse training samples that cover various formats and scenarios, improving data availability and model generalization while maintaining a manageable training complexity through controlled parameter variation.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3944145B1Method and device for training image recognition model, equipment and medium
Publication Date: 2024.03.20 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP3944145B1 patent drawingFigure 1~2
  • EP3944145B1 patent drawingFigure 3A~3B
  • EP3944145B1 patent drawingFigure 3C~4

AI summary

A method includes: acquiring training data, the training data includes training images for a preset vertical type, and the training images include a first training image containing real data of the preset vertical type and a second training image containing virtual data of the preset vertical type; building a basic model, the basic model includes a deep learning network, and the deep learning network is configured to recognize the training images to extract text data in the training image; and training the basic model by using the training data to obtain the image recognition model.