Face anonymization using a generative adversarial network

Patent metadata

Field	Value
Publication number	US-12438849-B2
Application number	US-202318131147-A
Country	US
Kind code	B2
Filing date	Apr 5, 2023
Priority date	Apr 6, 2022
Publication date	Oct 7, 2025
Grant date	Oct 7, 2025

Abstract

A generative adversarial network performs face anonymization. In a first step, an input image showing a face to be anonymized is received. Furthermore, an input vector with control data for face anonymization is received. The generative adversarial network then generates an output image in which the face is anonymized in accordance with the control data of the input vector.

First claim

The invention claimed is:

1. A non-transitory computer-readable medium having stored thereon computer-executable instructions that, when executed by a processor, perform operations face anonymization, the operations comprising:
receiving an input image showing a face to be anonymized;
receiving an input vector with control data for face anonymization; and
generating by a generative adversarial network, an output image in which the face is anonymized in accordance with the control data of the input vector, wherein the generative adversarial network comprises a generator sub-network trained to generate new faces with different facial attributes, a discriminator sub-network trained to evaluate if generated new faces are realistic and natural, and an identity classifier sub-network trained to evaluate if face identities of new faces have been changed, and wherein the generator sub-network, the discriminator sub-network and the identity classifier sub-network are trained alternately and compete with each other to achieve balance of optimization.

2. The non-transitory computer-readable medium according to claim 1 , wherein the input vector comprises control data for individual facial attributes.

3. The non-transitory computer-readable medium according to claim 2 , wherein the control data specifies whether a facial attribute shall be modified, kept, or optionally modified.

4. The non-transitory computer-readable medium according to claim 3 , wherein the control data further specifies an amount of modification of a facial attribute.

5. The non-transitory computer-readable medium according to claim 1 , wherein the generative adversarial network uses an adversarial loss function, a reconstruction loss function, an attribute loss function, and an identity loss function.

6. The non-transitory computer-readable medium according to claim 1 , wherein the input vector is selected based on an application scenario.

7. An apparatus for face anonymization, the apparatus comprising a generative adversarial network configured to:
receive an input image showing a face to be anonymized;
receive an input vector with control data for face anonymization; and
generate an output image in which the face is anonymized in accordance with the control data of the input vector, wherein the generative adversarial network comprises a generator sub-network trained to generate new faces with different facial attributes, a discriminator sub-network trained to evaluate if generated new faces are realistic and natural, and an identity classifier sub-network trained to evaluate if face identities of new faces have been changed, and wherein the generator sub-network, the discriminator sub-network, and the identity classifier sub-network are trained alternately and compete with each other to achieve balance of optimization.

8. The apparatus according to claim 7 , wherein the input vector comprises control data for individual facial attributes.

9. The apparatus according to claim 8 , wherein the control data specifies whether a facial attribute shall be modified, kept, or optionally modified.

10. The apparatus according to claim 9 , wherein the control data further specifies an amount of modification of a facial attribute.

11. The apparatus according to claim 7 , wherein the generative adversarial network uses an adversarial loss function, a reconstruction loss function, an attribute loss function, and an identity loss function.

12. The apparatus according to claim 7 , wherein the input vector is selected based on an application scenario.

Assignees

• Elektrobit Automotive Gmbh
• Univ Nanyang Tech

Inventors

• Wang Yi Estelle
• Zhai Liming
• Guo Qing
• Dahan Jeremy

Classifications

• G06T2207/30252

  Vehicle exterior; Vicinity of vehicle · CPC title

• G06T2207/30201

  Face · CPC title

• G06T2207/30168

  Image quality inspection · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T11/60

  Creating or editing images; Combining images with text · CPC title