	US 12,469,265 B2
	Method for neural network adaptation
Te Tang, Fremont, CA (US); and Tetsuaki Kato, Fremont, CA (US)
Assigned to FANUC CORPORATION, Yamanashi (JP)
Filed by FANUC CORPORATION, Yamanashi (JP)
Filed on Jan. 30, 2023, as Appl. No. 18/161,305.
Prior Publication US 2024/0257505 A1, Aug. 1, 2024
Int. Cl. G06V 10/776 (2022.01); B25J 9/16 (2006.01); G06V 10/26 (2022.01); G06V 10/764 (2022.01); G06V 10/77 (2022.01); G06V 10/774 (2022.01); G06V 10/82 (2022.01)

CPC G06V 10/776 (2022.01) [B25J 9/1697 (2013.01); G06V 10/26 (2022.01); G06V 10/764 (2022.01); G06V 10/7715 (2022.01); G06V 10/7747 (2022.01); G06V 10/82 (2022.01); G06V 2201/06 (2022.01)]

14 Claims

1. A system for adapting a feature extraction neural network, said system comprising:

a computer including at least one processor and a memory device storing data and executable code that, when executed, causes the at least one processor to:

train the feature extraction neural network using 2D red-green-blue (RGB) training images;

generate training features images by providing the training images to the trained feature extraction neural network;

provide the training features images to a dataset classifier neural network along with an identifier that the training features images are from the training images;

observe how accurately the dataset classifier neural network identified that the training features images were from the training images;

modify weights within the dataset classifier neural network to improve the accuracy of the ability of the dataset classifier neural network to identify that the training features images are from the training images;

generate test features images by providing test images to the trained feature extraction neural network, said test images including environment dependent features that are different than environment dependent features in the training images and environment independent features that are the same or similar to environment independent features in the training images;

provide the test features images to the dataset classifier neural network along with an identifier that the test features images are from the test images;

observe how accurately the dataset classifier neural network identified that the test features images were from the test images; and

modify the weights within the dataset classifier neural network to improve the accuracy of the ability of the dataset classifier neural network to identify that the test features images are from the test images;

and where the at least one processor;

again generates training features images by providing the training images to the trained feature extraction neural network;

again provides the training features images to the dataset classifier neural network, but without the identifier that the training features images are from the training images;

again observes how accurately the dataset classifier neural network identified that the training features images were from the training images;

modifies the weights within the feature extraction neural network to reduce the accuracy of the ability of the dataset classifier neural network to identify that the training features images are from the training images;

again generates test features images by providing the test images to the trained feature extraction neural network;

again provides the test features images to the dataset classifier neural network, but without the identifier that the test features images are from the test images;

again observes how accurately the dataset classifier neural network identified that the test features images were from the test images; and

modifies the weights within the feature extraction neural network to reduce the accuracy of the ability of the dataset classifier neural network to identify that the test features images are from the test images;

and wherein modifying the weights within the dataset classifier neural network and modifying the weights within the feature extraction neural network are repeated in an alternating sequence.