Face De-identification via Segmented Feature Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing techniques fail to effectively de-identify faces in images while maintaining similarity to human perception, thereby allowing face recognition algorithms to identify individuals.
Innovation Solution
The method employs selective image processing techniques using pre-provided face models to apply variations in geometry, features, and texture, optimizing the resemblance to the original face while reducing recognition scores below a certain threshold, utilizing geometry and feature de-identification processing, and adding noise layers to limit facial recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing image processing techniques are used to de-identify faces, then processing speed is maintained, but face recognition algorithms can still identify individuals
Solution Approach 1:
The patent segments the face image into multiple regions (eyes, nose, mouth, cheeks, forehead) and applies different processing techniques to each region. This segmentation allows the system to effectively de-identify the face by modifying specific features while maintaining overall facial structure, preventing recognition algorithms from identifying individuals while managing processing complexity through region-specific operations.
Solution Approach 2:
The patent applies local quality changes by modifying specific facial features (eyes, nose, mouth, cheeks, forehead) with different processing intensities and techniques. Each region receives tailored modifications that collectively de-identify the face while preserving natural appearance, addressing the contradiction between de-identification effectiveness and processing complexity.
2Reliability
If face features are significantly modified to prevent recognition, then de-identification effectiveness is improved, but resemblance to the original face is lost
Solution Approach 1:
The patent applies parameter changes by modifying facial features through controlled transformations that alter geometric and textural properties. By adjusting parameters such as feature positions, sizes, and textures within specific ranges, the system achieves effective de-identification while maintaining sufficient resemblance to the original face for human recognition.
Solution Approach 2:
The patent introduces asymmetry by applying non-uniform modifications to symmetric facial features. This asymmetric processing disrupts the precise symmetry that recognition algorithms rely upon, improving de-identification effectiveness while maintaining enough structural similarity for human perception of facial resemblance.
3Reliability
If multiple processing techniques are applied to de-identify faces, then de-identification effectiveness is improved, but processing time increases
Solution Approach 1:
The patent applies preliminary action by performing face detection and feature localization before applying de-identification processing. This preliminary segmentation and identification of key facial regions enables subsequent processing to be targeted and efficient, reducing overall processing time while maintaining effective de-identification through focused modifications to identified features.
Solution Approach 2:
The patent applies partial action by selectively processing only the most critical facial features (eyes, nose, mouth, cheeks, forehead) rather than the entire face image. This selective approach achieves effective de-identification by modifying key recognition features while minimizing processing time through targeted operations on essential regions only.
Data Source
AI summary
System and method for training a human perception predictor to determine level of perceived similarity between data samples, the method including: receiving at least one media file, determining at least one identification region for each media file, applying at least one transformation on each identification region for each media file until at least one modified media file is created, receiving input regarding similarity between each modified media file and the corresponding received media file, and training a machine learning model with an objective function configured to predict similarity between media files by a human observer in accordance with the received input.