Facial Image Anonymization Preserving Expression Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for developing automatic facial expression recognition systems face challenges due to the lack of diverse and labeled facial images, often resulting in overfitting when using images of similar-looking individuals, and the labor-intensive process of labeling images for training, which can be mitigated by anonymizing facial images while preserving emotional expressions.
Innovation Solution
The method involves encoding original facial images into feature sets that separate personal identity and expression components, applying a perturbation transform to obscure personal identity while preserving expression components, and decoding these feature sets to generate anonymized images that can be used for training facial expression recognition systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If facial images are anonymized to protect subject identity, then privacy protection is improved, but facial expression information may be lost
Solution Approach 1:
The patent segments the facial image into distinct feature components: identity features (e.g., facial structure, geometry) and expression features (e.g., muscle movements, emotional cues). The anonymization process selectively pertains only the identity features while leaving expression features intact, allowing the system to differentiate between these two types of information and apply differential processing accordingly.
Solution Approach 2:
The patent applies local quality by applying different levels of perturbation to different regions of the facial image. Identity-critical regions (such as facial contours and bone structure) receive strong perturbation to ensure anonymization, while expression-relevant regions (such as muscle movements and facial gestures) receive minimal or no perturbation to preserve expression information for training purposes.
2Adaptability or versatility
If a large diverse training set is collected from the internet, then training data diversity is improved, but labeling labor intensity increases
Solution Approach 1:
The patent enables self-service by using the anonymization system to automatically generate diverse training data from a small set of original images. The system applies perturbation transforms to create multiple anonymized variations of each original image, effectively multiplying the training data volume without requiring proportional increases in human labeling effort, as the anonymized images inherit labels from the original images.
Solution Approach 2:
The patent uses copying by creating multiple synthetic copies of the original facial images through perturbation transforms. Each copy is a modified version that preserves expression information while altering identity features, allowing the system to generate a large diverse training set by copying and transforming a small number of original images rather than collecting many unique images.
3Ease of manufacture
If images of similar-looking people are used for training, then data collection is simplified, but model overfitting increases
Solution Approach 1:
The patent applies parameter changes by systematically varying multiple parameters of the facial images through perturbation transforms. These parameters include but are not limited to: lighting conditions, image quality, facial expression intensity, and anonymization strength. By changing these parameters across multiple generated images, the system creates diverse training data that prevents overfitting while maintaining ease of data collection from a small original set.
Data Source
AI summary
A method facilitates the use of facial images through anonymization of facial images, thereby allowing people to submit their own facial images without divulging their identities. Original facial images are accessed and perturbed to generate synthesized facial images. Personal identities contained in the original facial images are no longer discernable from the synthesized facial images. At the same time, each synthesized facial image preserves at least some of the original attributes of the corresponding original facial image.