US 12,148,194 B2
Method, device, and storage medium for targeted adversarial discriminative domain adaptation
Hua-mei Chen, Germantown, MD (US); Ashley Diehl, Dayton, OH (US); Erik Blasch, Arlington, VA (US); and Genshe Chen, Germantown, MD (US)
Assigned to Intelligent Fusion Technology, Inc., Germantown, MD (US)
Filed by Intelligent Fusion Technology, Inc., Germantown, MD (US)
Filed on Sep. 14, 2021, as Appl. No. 17/474,516.
Claims priority of provisional application 63/080,291, filed on Sep. 18, 2020.
Claims priority of provisional application 63/078,073, filed on Sep. 14, 2020.
Prior Publication US 2024/0185555 A1, Jun. 6, 2024
Int. Cl. G06V 10/44 (2022.01); G06V 10/764 (2022.01)
CPC G06V 10/44 (2022.01) [G06V 10/764 (2022.01)] 15 Claims
OG exemplary drawing
 
1. A targeted adversarial discriminative domain adaptation (T-ADDA) method, comprising:
pre-training a source model, including a source feature encoder and a source classifier, on a source domain image dataset according to combined cross-entropy loss and center loss functions, wherein source feature vectors in each source class are generated, the center loss function is expressed by

OG Complex Work Unit Math
 xi and yi denote an i-th source feature vector and a label corresponding to the i-th source feature vector, respectively, Cyi denotes a vi-th source feature class center, and m is a number of labeled source images;
adapting a target feature encoder by:
configuring the pre-trained source feature encoder of the pre-trained source model to be an initial target feature encoder, and using the initial target feature encoder to generate target feature vectors in each target class based on a target domain image dataset;
adjusting a domain discriminator according to an adversarial domain discrimination loss using the source feature vectors in each source class and the target feature vectors in each target class;
adjusting the initial target feature encoder according to a generative adversarial network (GAN) loss using the target feature vectors in each target class and the adjusted domain discriminator; and
further adjusting the initial target feature encoder to generate the target feature encoder according to a feature class matching loss using labeled target feature vectors and corresponding source feature class centers; and
generating a target model by concatenating the adapted target feature encoder with the pre-trained source classifier of the pre-trained source model.