Selective Character Anonymization for OCR Training Images

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for anonymizing image data fail to effectively remove sensitive personal information while preserving the image's utility for training systems, particularly for optical character recognition, and often degrade the training effectiveness or render the data unusable.

Innovation Solution

A computer-implemented method that selectively modifies image data by removing or blurring sensitive information, such as alphanumeric characters, faces, and machine-readable codes, while maintaining structural and imaging properties, using techniques like character shuffling and replacement with unmodified sections from the same image.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If sensitive regions are anonymized using noise or geometric transformations, then personal information is protected, but the image data becomes useless for training

Engineering Contradiction:

Improvepersonal information protectionVSAvoidtraining effectiveness

Core Design Contradiction:

Object-affected harmful factorsVSReliability

Solution Approach 1:

The patent segments the text region into individual character units and selectively anonymizes only certain characters (e.g., every second character or specific positions) rather than anonymizing the entire text region. This partial segmentation approach protects personal information while preserving enough textual structure for training purposes.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies different anonymization treatments to different local regions within the text. Specifically, it anonymizes only certain character positions while leaving others unanonymized, creating a local quality variation that balances privacy protection with training utility. The anonymization is localized to specific character sections rather than applied uniformly.

Inventive Principle:

Principle #3Local quality

2Object-affected harmful factors

If text regions are anonymized by replacing with generic image parts, then personal information is removed, but structural information is destroyed

Engineering Contradiction:

Improvesensitive information removalVSAvoidstructural information

Core Design Contradiction:

Object-affected harmful factorsVSLoss of information

Solution Approach 1:

The patent copies unanonymized character sections from the same image to serve as replacement elements for anonymized characters. This copying approach maintains the original structural information, font characteristics, and visual style while still implementing anonymization through selective removal or substitution of specific character sections.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

Instead of replacing anonymized characters with generic external image parts (as in prior art), the patent inverts the approach by using character sections from the same image as replacements. This inversion preserves the internal structural consistency and eliminates the need for external generic parts that would destroy structural information.

Inventive Principle:

Principle #13The other way round (Inversion)

3Object-affected harmful factors

If all alphanumeric characters are anonymized, then personal information is protected, but training data quality deteriorates

Engineering Contradiction:

Improvepersonal information protectionVSAvoidtraining data quality

Core Design Contradiction:

Object-affected harmful factorsVSManufacturing precision

Solution Approach 1:

The patent applies partial anonymization by selectively anonymizing only certain alphanumeric characters (e.g., 50% of characters or specific positions) rather than anonymizing all characters. This partial action approach provides sufficient privacy protection while maintaining enough recognizable text patterns for effective training of optical character recognition systems.

Inventive Principle:

Principle #16Partial or excessive action

Solution Approach 2:

The patent changes the parameter of anonymization intensity from 100% (all characters) to a controlled percentage (e.g., 50% or specific positions). This parameter adjustment allows optimization between privacy protection and training quality, enabling the system to learn from partially anonymized text while still protecting personal information.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12443751B2Anonymizing textual content in image training data

Publication Date: 2025.10.14 ANYLINE GMBH

AI summary

A computer-implemented method for modifying image data, the method including: loading unmodified image data; detecting at least two alphanumeric characters in an image represented by the loaded image data; selecting one or more of the detected alphanumeric characters, wherein the number of selected alphanumeric characters is smaller than the total number of detected alphanumeric characters; modifying the loaded image data by removing one or more character sections of the loaded image data, wherein each character section corresponds to an area of a selected alphanumeric character; storing the modified image data. Also provided is a corresponding data processing system, computer program product and computer-readable storage medium.