Deep image delighting

Patent metadata

Field	Value
Publication number	US-2025238898-A1
Application number	US-202418419942-A
Country	US
Kind code	A1
Filing date	Jan 23, 2024
Priority date	Jan 23, 2024
Publication date	Jul 24, 2025
Grant date	—

Abstract

A method for content modification executed by a computer including a processor and a memory includes receiving an initial image that includes light emitted from a representation of a light source. The method further includes analyzing the initial image, with a delighting model, to predict a second image comprising negative light of the initial image. The method further includes subtracting pixel values of the predicted second image comprising negative light of the initial image from pixel values of the initial image. The method further includes outputting a third image comprising the initial image without the emitted light from the representation of the light source. A second method includes training a delighting model with the initial image with an added predetermined light source and the predetermined light source, to predict an image comprising negative light of the initial image.

First claim

1 . A method for content modification executed by a computer including a processor and a memory comprising:
receiving an initial image wherein the initial image includes light emitted from a representation of a light source; analyzing the initial image, with a delighting model, to predict a second image comprising the light emitted from the representation of the light source; subtracting pixel values of the predicted second image comprising the light emitted from the representation of the light source from pixel values of the initial image; and outputting a third image comprising the initial image without the emitted light from the representation of the light source.

2 . The method of claim 1 , wherein the initial image includes light emitted from a one or more representations of light sources.

3 . The method of claim 1 , further comprising identifying a perimeter shape of the representation of the light source in the initial image.

4 . The method of claim 1 , wherein the delighting model is trained on training data comprising a predetermined light source and an image with the predetermined light source added to it.

5 . The method of claim 1 , wherein the delighting model predicts the second image based on an identified perimeter shape of the representation of the light source and the initial image as inputs.

6 . The method of claim 1 , further comprising:
modifying the initial image to remove a housing of the representation of the light source, leaving light emitted from the representation of the light source; and using the modified initial image as an input in the delighting model in place of the initial image.

7 . The method of claim 1 , further comprising using the delighting model to predict negative light of one or more stray light sources within the initial image including a halo effect.

8 . The method of claim 1 , further comprising using the delighting model to predict negative light of one or more lens effects within the initial image including a lens flare.

9 . The method of claim 1 further comprising using the delighting model to add one or more new representation of light sources and corresponding emitted light from the one or more new representations of the light sources to the outputted third image.

10 . The method of claim 9 further comprising using the delighting model to generate an animated image with one or more animated representations of light sources.

11 . The method of claim 1 , further comprising identifying the representation of the light source within an initial image comprising one or more light sources.

12 . A method for content modification executed by a computer including a processor and a memory comprising:
receiving an initial image; adding, to the initial image, a predetermined light source; training a delighting model with the initial image including the added predetermined light source and the predetermined light source, to predict a negative light image comprising only light emitted from a light source; and outputting the trained delighting model.

13 . The method of claim 12 , wherein the delighting model is trained with one or more initial images.

14 . The method of claim 12 , wherein one or more predetermined light sources are added to the initial image.

15 . The method of claim 12 , wherein the predetermined light source comprises a segmentation of the predetermined light source.

16 . The method of claim 12 , wherein the initial image comprises background lighting effects.

17 . The method of claim 12 , wherein the predetermined light source comprises one or more lens effects including a lens flare.

18 . The method of claim 12 , wherein the predetermined light source comprises one or more halo effects.

19 . The method of claim 12 , wherein adding the predetermined light source to the initial image occurs during training of the delighting model.

20 . The method of claim 12 , further comprising identifying the shape and location of the predetermined light source using a segmentation model.

Assignees

• Disney Entpr Inc

Inventors

• Iglesias Navarro Santiago
• Pernías Pascual De Pobil Pablo
• Moore Robert B
• Juboor David N

Classifications

• G06T5/60

  using machine learning, e.g. neural networks · CPC title

• G06T2207/20224

  Image subtraction · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T5/94 Primary

  based on local image properties, e.g. for local contrast enhancement · CPC title

• G06T5/77

  Retouching; Inpainting; Scratch removal · CPC title