CPC H04N 7/157 (2013.01) [G06F 3/013 (2013.01); G06N 3/04 (2013.01); G06N 3/045 (2023.01); G06T 7/11 (2017.01); G06T 7/70 (2017.01); G06T 15/04 (2013.01); G06T 15/20 (2013.01); G06T 15/205 (2013.01); G06T 17/20 (2013.01); G06T 19/00 (2013.01); G06T 19/20 (2013.01); H04N 7/144 (2013.01); H04N 7/147 (2013.01); H04N 7/152 (2013.01); G06T 2200/08 (2013.01); G06T 2207/30201 (2013.01); G06T 2219/2004 (2013.01)] | 20 Claims |
1. A method for generating alpha channel information related to a person, the method comprises:
generating the alpha channel information related to the person by an alpha channel neural network; wherein the alpha channel machine learning process is trained using ground truth alpha channel information that is associated with background information;
wherein the ground-truth alpha channel information is generated by:
obtaining input images of the person; wherein each image comprises the person and an arbitrary background;
repeating, for each input image of the input images, converting the input image, by a portrait matting predictor, to (a) a first colored background image that comprises the person and a background of a first color, and (b) a second colored background image that comprises the person and a background of a second color;
wherein the first color differs from the second color;
wherein the training comprises:
repeating, for each input image of the input images:
i. determining person properties within the input image;
ii. generating, by the alpha channel neural network, a first image, the first image is of the first avatar of the person with the first colored background and having the person properties;
iii. comparing first alpha channel information related to the first image to ground-truth alpha channel information related to the first colored background image to provide a first comparison result;
iv. generating, by the alpha channel neural network, a second image, the second image is of the second avatar of the person with the second colored background and having the person properties; and
v. comparing second alpha channel information related to the second image to ground-truth alpha channel information related to the second colored background image to provide a second comparison result.
|