Face De-Identification Model Using Virtual Identity Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image de-identification methods using conditional generative adversarial networks (GANs) result in poor-quality anonymized images due to the use of background pictures with face areas removed, leading to loss of identity-free attributes.
Innovation Solution
A model training method that projects training images to a target space to obtain virtual identity vectors and extracts attribute vectors, generating de-identified images based on these vectors while ensuring attribute information remains unchanged, using a projector, attribute extractor, and generator in the target network model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If background picture with face area removed is used as model input, then identity feature removal is achieved, but image quality deteriorates
Solution Approach 1:
The patent extracts only the necessary attribute information (pose, scale, background) from the original image, rather than removing the entire face area. This selective extraction preserves useful information while achieving de-identification, resolving the contradiction between identity removal and quality preservation.
Solution Approach 2:
The patent transforms the input representation from removed face areas to attribute vectors (pose angles, scale factors, background features). By changing the parameter representation, the model receives condensed essential information without identity features, maintaining quality while achieving de-identification.
2Reliability
If face area is removed from original picture, then de-identification is achieved, but attribute information is lost
Solution Approach 1:
The patent extracts specific attribute information (pose, scale, background) separately from the face area, preserving these attributes while removing only the identity-containing face region. This selective extraction prevents information loss of non-identity attributes.
Solution Approach 2:
The patent performs preliminary extraction of attribute information before face removal. By pre-extracting pose, scale, and background attributes, the system preserves this information for later reconstruction, preventing attribute loss during the de-identification process.
3Manufacturing precision
If virtual identity vectors and attribute vectors are used for image generation, then image quality is improved, but model complexity increases
Solution Approach 1:
The patent segments the generation process into distinct components: virtual identity vector generation, attribute vector extraction, and image synthesis. This segmentation allows each component to be optimized independently, managing complexity while improving overall image quality through specialized processing for each aspect.
Data Source
AI summary
An de-identification method includes: performing sampling on a target space of a projector in a target network model to obtain N virtual identity vectors, N being a positive integer; extracting, by using an attribute extractor in the target network model, M attribute vectors from a target image, M being a positive integer; and generating, by using a generator in the target network model, a de-identified image of the target image based on the N virtual identity vectors and the M attribute vectors.