US 12,327,188 B2
Direct regression encoder architecture and training
Ratheesh Kalarot, San Jose, CA (US); Wei-An Lin, San Jose, CA (US); Cameron Smith, Santa Cruz, CA (US); Zhixin Shu, San Jose, CA (US); Baldo Faieta, San Francisco, CA (US); Shabnam Ghadar, Menlo Park, CA (US); Jingwan Lu, Santa Clara, CA (US); Aliakbar Darabi, Seattle, WA (US); Jun-Yan Zhu, San Jose, CA (US); Niloy Mitra, London (GB); Richard Zhang, San Francisco, CA (US); and Elya Shechtman, Seattle, WA (US)
Assigned to Adobe Inc., San Jose, CA (US)
Filed by Adobe Inc., San Jose, CA (US)
Filed on Jul. 23, 2021, as Appl. No. 17/384,371.
Claims priority of provisional application 63/092,980, filed on Oct. 16, 2020.
Prior Publication US 2022/0121931 A1, Apr. 21, 2022
Int. Cl. G06F 3/0484 (2022.01); G06F 3/04845 (2022.01); G06F 3/04847 (2022.01); G06F 18/21 (2023.01); G06F 18/211 (2023.01); G06F 18/214 (2023.01); G06F 18/40 (2023.01); G06N 3/045 (2023.01); G06N 3/08 (2023.01); G06N 20/20 (2019.01); G06T 3/02 (2024.01); G06T 3/18 (2024.01); G06T 3/40 (2006.01); G06T 3/4038 (2024.01); G06T 3/4046 (2024.01); G06T 5/20 (2006.01); G06T 5/77 (2024.01); G06T 11/00 (2006.01); G06T 11/60 (2006.01); G06V 10/28 (2022.01); G06V 10/98 (2022.01)
CPC G06N 3/08 (2013.01) [G06F 3/04845 (2013.01); G06F 3/04847 (2013.01); G06F 18/211 (2023.01); G06F 18/214 (2023.01); G06F 18/2163 (2023.01); G06F 18/40 (2023.01); G06N 3/045 (2023.01); G06N 20/20 (2019.01); G06T 3/02 (2024.01); G06T 3/18 (2024.01); G06T 3/40 (2013.01); G06T 3/4038 (2013.01); G06T 3/4046 (2013.01); G06T 5/20 (2013.01); G06T 5/77 (2024.01); G06T 11/001 (2013.01); G06T 11/60 (2013.01); G06V 10/28 (2022.01); G06V 10/98 (2022.01); G06T 2207/10024 (2013.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/20221 (2013.01); G06T 2210/22 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
providing an input image as input to an encoder neural network, the encoder neural network implemented by one or more computing systems, the encoder neural network comprising an input layer, a feature extraction layer, a bottleneck layer positioned after the feature extraction layer, and a series of fully connected layers positioned after the bottleneck layer;
producing, by the encoder neural network, a latent space representation of the input image, wherein the producing comprises extracting a feature vector from the feature extraction layer, providing the feature vector to the bottleneck layer as input to generate a feature vector of reduced dimensionality, providing the feature vector of reduced dimensionality as input to the series of fully connected layers to produce a latent space representation of dimensionality greater than the output of the bottleneck layer and less than the output of the feature extraction layer, and producing the latent space representation as output;
providing the latent space representation produced by the encoder neural network as input to a generator neural network, the generator neural network implemented using the one or more computing systems; and
generating, by the generator neural network, an output image based upon the latent space representation.