Granted Patent B2

US 12,307,545 · App. 17/985,578 · Granted May 20, 2025

Inventors: Zijia Wang (WeiFang, CN); Zhisong Liu (Shenzhen, CN); Zhen Jia (Shanghai, CN)

Assignee: Dell Products L.P.

G06T1/0021G06T13/40G06T2207/20081

Abstract

Claims (88)

1. A method of processing a virtual avatar, comprising:

generating an image feature of the virtual avatar based on a plurality of image blocks of the virtual avatar and corresponding positions of the plurality of image blocks in the virtual avatar;

generating, based on a watermark to be added to the virtual avatar, a text feature associated with text of the watermark; and

generating a watermarked virtual avatar based on the image feature and the text feature, wherein the watermark is invisible to human beings and identifies an identity of a user of the virtual avatar in a metaverse;

wherein generating the watermarked virtual avatar based on the image feature and the text feature comprises:

determining a key-value pair based on the image feature and the text feature, the key-value pair comprising (i) a key that represents a semantic embedding parameter of a self-attention mechanism in a machine learning model, and (ii) a value that represents the text feature;

determining the image feature as a query vector;

determining a weight set based on a similarity between the query vector and a key in the key-value pair; and

determining, based on the weight set, a watermarked image feature;

wherein generating the watermarked virtual avatar further comprises processing the watermarked virtual avatar through an invariant layer of the machine learning model, the invariant layer comprising at least one fully-connected layer of the machine learning model and implementing a mapping function that converts a first number of image channels of an input image space of the watermarked virtual avatar into a second number of image channels of a transformed image space of an updated watermarked virtual avatar, the second number being greater than the first number.

2. The method according to claim 1 , wherein generating the image feature of the virtual avatar comprises:

dividing the virtual avatar into a plurality of visual marks, wherein each visual mark in the plurality of visual marks is represented by a vector of a predetermined length;

generating a corresponding position code for each of the visual marks; and

generating the image feature based on the plurality of visual marks and the corresponding position codes.

3. The method according to claim 1 , wherein generating the text feature associated with the text of the watermark comprises:

dividing the watermark into a plurality of text lemmas, wherein each text lemma in the plurality of text lemmas comprises a word; and

generating the text feature based on the plurality of text lemmas.

4. The method according to claim 1 , wherein the key-value pair is a first key-value pair, the query vector is a first query vector, the weight set is a first weight set, and the method further comprises:

determining a second key-value pair based on the watermarked image feature and the text feature;

determining the watermarked image feature as a second query vector;

determining a second weight set based on a similarity between the second query vector and a key in the second key-value pair; and

determining, based on the second weight set, an updated watermarked image feature.

5. The method according to claim 4 , wherein generating the watermarked virtual avatar based on the image feature and the text feature comprises:

generating the watermarked virtual avatar based on the updated watermarked image feature and the text feature.

6. The method according to claim 1 , further comprising:

sending the updated watermarked virtual avatar to a metaverse platform; and

sending an instruction for extracting the watermark in the watermarked virtual avatar to the metaverse platform.

7. The method according to claim 6 , further comprising:

causing, in response to the watermark being extracted, the metaverse platform to verify the watermark based on information of the user; and

causing, based on a result of the verification of the watermark, the metaverse platform to determine at least one of authenticity and ownership of the virtual avatar.

8. The method according to claim 1 , wherein the method is performed by a trained machine learning model, and the method further comprises:

training the machine learning model by using sample data, wherein the sample data comprises the virtual avatar, the watermark and a corresponding labeled watermarked virtual avatar, and a labeled extracted watermark.

9. An electronic device, comprising:

a processor; and

a memory coupled to the processor, wherein the memory has instructions stored therein which, when executed by the processor, cause the electronic device to execute actions comprising:

generating an image feature of a virtual avatar based on a plurality of image blocks of the virtual avatar and corresponding positions of the plurality of image blocks in the virtual avatar;

generating, based on a watermark to be added to the virtual avatar, a text feature associated with text of the watermark; and

generating a watermarked virtual avatar based on the image feature and the text feature, wherein the watermark is invisible to human beings and identifies an identity of a user of the virtual avatar in a metaverse;

wherein generating the watermarked virtual avatar based on the image feature and the text feature comprises:

determining a key-value pair based on the image feature and the text feature, the key-value pair comprising (i) a key that represents a semantic embedding parameter of a self-attention mechanism in a machine learning model, and (ii) a value that represents the text feature;

determining the image feature as a query vector;

determining a weight set based on a similarity between the query vector and a key in the key-value pair; and

determining, based on the weight set, a watermarked image feature;

wherein generating the watermarked virtual avatar further comprises processing the watermarked virtual avatar through an invariant layer of the machine learning model, the invariant layer comprising at least one fully-connected layer of the machine learning model and implementing a mapping function that converts a first number of image channels of an input image space of the watermarked virtual avatar into a second number of image channels of a transformed image space of an updated watermarked virtual avatar, the second number being greater than the first number.

10. The electronic device according to claim 9 , wherein generating the image feature of the virtual avatar comprises:

dividing the virtual avatar into a plurality of visual marks, wherein each visual mark in the plurality of visual marks is represented by a vector of a predetermined length;

generating a corresponding position code for each of the visual marks; and

generating the image feature based on the plurality of visual marks and the corresponding position codes.

11. The electronic device according to claim 9 , wherein generating the text feature associated with the text of the watermark comprises:

dividing the watermark into a plurality of text lemmas, wherein each text lemma in the plurality of text lemmas comprises a word; and

generating the text feature based on the plurality of text lemmas.

12. The electronic device according to claim 9 , wherein the key-value pair is a first key-value pair, the query vector is a first query vector, the weight set is a first weight set, and the actions further comprise:

determining a second key-value pair based on the watermarked image feature and the text feature;

determining the watermarked image feature as a second query vector;

determining a second weight set based on a similarity between the second query vector and a key in the second key-value pair; and

determining, based on the second weight set, an updated watermarked image feature.

13. The electronic device according to claim 12 , wherein generating the watermarked virtual avatar based on the image feature and the text feature comprises:

generating the watermarked virtual avatar based on the updated watermarked image feature and the text feature.

14. The electronic device according to claim 9 , wherein the actions further comprise:

sending the updated watermarked virtual avatar to a metaverse platform; and

sending an instruction for extracting the watermark in the watermarked virtual avatar to the metaverse platform.

15. The electronic device according to claim 14 , wherein the actions further comprise:

causing, in response to the watermark being extracted, the metaverse platform to verify the watermark based on information of the user; and

causing, based on a result of the verification of the watermark, the metaverse platform to determine at least one of authenticity and ownership of the virtual avatar.

16. A computer program product tangibly stored on a non-transitory computer-readable medium and comprising computer-executable instructions, wherein the computer-executable instructions, when executed by a device, cause the device to implement a method of processing a virtual avatar, the method comprising:

generating an image feature of the virtual avatar based on a plurality of image blocks of the virtual avatar and corresponding positions of the plurality of image blocks in the virtual avatar;

generating, based on a watermark to be added to the virtual avatar, a text feature associated with text of the watermark; and

generating a watermarked virtual avatar based on the image feature and the text feature, wherein the watermark is invisible to human beings and identifies an identity of a user of the virtual avatar in a metaverse;

wherein generating the watermarked virtual avatar based on the image feature and the text feature comprises:

determining a key-value pair based on the image feature and the text feature, the key-value pair comprising (i) a key that represents a semantic embedding parameter of a self-attention mechanism in a machine learning model, and (ii) a value that represents the text feature;

determining the image feature as a query vector;

determining a weight set based on a similarity between the query vector and a key in the key-value pair; and

determining, based on the weight set, a watermarked image feature;

wherein generating the watermarked virtual avatar further comprises processing the watermarked virtual avatar through an invariant layer of the machine learning model, the invariant layer comprising at least one fully-connected layer of the machine learning model and implementing a mapping function that converts a first number of image channels of an input image space of the watermarked virtual avatar into a second number of image channels of a transformed image space of an updated watermarked virtual avatar, the second number being greater than the first number.

17. The computer program product of claim 16 , wherein generating the image feature of the virtual avatar comprises:

dividing the virtual avatar into a plurality of visual marks, wherein each visual mark in the plurality of visual marks is represented by a vector of a predetermined length;

generating a corresponding position code for each of the visual marks; and

generating the image feature based on the plurality of visual marks and the corresponding position codes.

18. The computer program product of claim 16 , wherein generating the text feature associated with the text of the watermark comprises:

dividing the watermark into a plurality of text lemmas, wherein each text lemma in the plurality of text lemmas comprises a word; and

generating the text feature based on the plurality of text lemmas.

19. The computer program product of claim 16 , wherein the key-value pair is a first key-value pair, the query vector is a first query vector, the weight set is a first weight set, and the method further comprises:

determining a second key-value pair based on the watermarked image feature and the text feature;

determining the watermarked image feature as a second query vector;

determining a second weight set based on a similarity between the second query vector and a key in the second key-value pair; and

determining, based on the second weight set, an updated watermarked image feature.

20. The computer program product of claim 19 , wherein generating the watermarked virtual avatar based on the image feature and the text feature comprises:

generating the watermarked virtual avatar based on the updated watermarked image feature and the text feature.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (18)