Photo relighting and background replacement based on machine learning models

Patent metadata

Field	Value
Publication number	US-12530744-B2
Application number	US-202118554960-A
Country	US
Kind code	B2
Filing date	Apr 28, 2021
Priority date	Apr 28, 2021
Publication date	Jan 20, 2026
Grant date	Jan 20, 2026

Abstract

Apparatus and methods related to applying lighting models to images are provided. An example method includes receiving, via a computing device, an image comprising a subject. The method further includes relighting, via a neural network, a foreground of the image to maintain a consistent lighting of the foreground with a target illumination. The relighting is based on a per-pixel light representation indicative of a surface geometry of the foreground. The light representation includes a specular component, and a diffuse component, of surface reflection. The method additionally includes predicting, via the neural network, an output image comprising the subject in the relit foreground. One or more neural networks can be trained to perform one or more of the aforementioned aspects.

First claim

What is claimed is:

1 . A computer-implemented method, comprising:
receiving, by a computing device, an image comprising a subject; relighting, by a neural network, a foreground of the image to maintain a consistent lighting of the foreground with a target illumination, wherein the relighting is based on a per- pixel light representation indicative of a surface geometry of the foreground, and wherein the light representation comprises a specular component, and a diffuse component, of surface reflection, and wherein the relighting comprises:
pre-filtering the target illumination to generate a plurality of candidate specular light maps,
predicting, by a specular net of the neural network, a per-pixel weight image based on the plurality of candidate specular light maps, and
generating a blended specular light map by taking a weighted sum of the plurality of candidate specular light maps using the per-pixel weight image; and
predicting, by the neural network, an output image comprising the subject in a relit foreground.

2 . The computer-implemented method of claim 1 , further comprising:
predicting, via the neural network, a reflectance field indicative of an appearance of the subject.

3 . The computer-implemented method of claim 1 , further comprising:
predicting, via the neural network and based on the foreground and the surface geometry, a per-pixel albedo image.

4 . The computer-implemented method of claim 1 , further comprising:
pre-filtering the target illumination, wherein the pre-filtering is based on a convolved specular light map and a convolved diffuse light map, and wherein the specular component and the diffuse component of the surface reflection are determined by sampling the convolved specular light map and the convolved diffuse light map by using one or more of a surface normal or a reflection vector.

5 . The computer-implemented method of claim 1 , further comprising:
receiving, as input to the neural network, the plurality of candidate light maps, an albedo, and an initial foreground; and predicting, via the neural network, the specular component of the per-pixel lighting representation.

6 . The computer-implemented method of claim 5 , wherein the relighting of the foreground further comprises:
concatenating the specular component, the diffuse component, and the albedo; and predicting, via the neural network, the relighting of the foreground based on the concatenating.

7 . The computer-implemented method of claim 1 , further comprising:
estimating, via the neural network, the foreground of the image.

8 . The computer-implemented method of claim 7 , wherein the estimating of the foreground comprises estimating an alpha matte.

9 . The computer-implemented method of claim 1 , wherein the relighting of the foreground comprises inferring one or more of a low-frequency color or a shading under the target illumination.

10 . The computer-implemented method of claim 1 , further comprising:
predicting, via the neural network, a plurality of per-pixel surface normal representations.

11 . The computer-implemented method of claim 1 , wherein an initial illumination associated with the image is different from the target illumination.

12 . The computer-implemented method of claim 1 , wherein the target illumination is not based on controlled lighting.

13 . The computer-implemented method of claim 1 , further comprising:
identifying a target background, wherein the target illumination is associated with the target background, and wherein the predicting of the output image comprises compositing the relit foreground into the target background using an alpha matte.

14 . The computer-implemented method of claim 13 , wherein the image comprises the subject in an initial background that is different from the target background.

15 . The computer-implemented method of claim 13 , further comprising:
receiving a second input image comprising a second subject in a second initial background; relighting a second foreground of the second image based on the target illumination; and compositing the relit foreground, the second relit foreground and the target background, and wherein the output image comprises the subject and the second subject in the target background.

16 . The computer-implemented method of claim 1 , further comprising:
training the neural network to receive a particular input image comprising a particular subject, and predict a particular output image comprising the particular subject in a relit foreground.

17 . The computer-implemented method of claim 16 , wherein the training of the neural network comprises training a ground truth alpha matte generation model.

18 . The computer-implemented method of claim 1 , wherein a training dataset comprises a plurality of images of subjects captured in a light stage computational illumination system, wherein the computational illumination system records one or more of: (i) a plurality of lighting conditions, (ii) a plurality of reflectance fields indicative of surface reflectance and surface geometry, or (iii) a plurality of alpha mattes.

19 . A computing device, comprising:
one or more processors; and data storage, wherein the data storage has stored thereon computer-executable instructions that, when executed by the one or more processors, cause the computing device to carry out functions comprising:
receiving, by the computing device, an image comprising a subject;
relighting, by a neural network, a foreground of the image to maintain a consistent lighting of the foreground with a target illumination, wherein the relighting is based on a per-pixel light representation indicative of a surface geometry of the foreground, and wherein the light representation comprises a specular component, and a diffuse component, of surface reflection, and wherein the relighting comprises:
pre-filtering the target illumination to generate a plurality of candidate specular light maps,
predicting, by a specular net of the neural network, a per-pixel weight image based on the plurality of candidate specular light maps, and
generating a blended specular light map by taking a weighted sum of the plurality of candidate specular light maps using the per-pixel weight image; and

predicting, by the neural network, an output image comprising the subject in a relit foreground.

20 . An article of manufacture comprising one or more non-transitory computer readable media having computer-readable instructions stored thereon that, when executed by one or more processors of a computing device, cause the computing device to carry out functions comprising:
receiving, by the computing device, an image comprising a subject; relighting, by a neural network, a foreground of the image to maintain a consistent lighting of the foreground with a target illumination, wherein the relighting is based on a per-pixel light representation indicative of a surface geometry of the foreground, and wherein the light representation comprises a specular component, and a diffuse component, of surface reflection, and wherein the relighting comprises:
pre-filtering the target illumination to generate a plurality of candidate specular light maps,
predicting, by a specular net of the neural network, a per-pixel weight image based on the plurality of candidate specular light maps, and
generating a blended specular light map by taking a weighted sum of the plurality of candidate specular light maps using the per-pixel weight image; and
predict

Assignees

• Google Llc

Inventors

• Legendre Chloe
• Debevec Paul
• Fanello Sean Ryan Francesco
• Pandey Rohit Kumar
• Escolano Sergio Orts
• Haene Christian
• Bouaziz Sofien

Classifications

• H04N5/272

  Means for inserting a foreground image in a background image, i.e. inlay, outlay · CPC title

• G06T2207/20221

  Image fusion; Image merging · CPC title

• G06V10/60

  relating to illumination properties, e.g. using a reflectance or lighting model · CPC title

• G06V10/56

  relating to colour · CPC title

• G06T7/194

  involving foreground-background segmentation · CPC title