US 12,437,581 B2
Training system, training method and recognition system
Chien-Hao Chen, Hsinchu (TW); Chih-Wei Wu, Hsinchu (TW); and Shih-Tse Chen, Hsinchu (TW)
Assigned to REALTEK SEMICONDUCTOR CORP., Hsinchu (TW)
Filed by REALTEK SEMICONDUCTOR CORP., Hsinchu (TW)
Filed on Aug. 21, 2023, as Appl. No. 18/236,142.
Claims priority of application No. 112105064 (TW), filed on Feb. 13, 2023.
Prior Publication US 2024/0273944 A1, Aug. 15, 2024
Int. Cl. G06V 40/16 (2022.01); G06V 10/82 (2022.01)
CPC G06V 40/172 (2022.01) [G06V 10/82 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A training system, comprising:
at least one processing unit and a neural network module, wherein the neural network module comprises:
an encoder module having a plurality of first parameters and configured to receive an input image and output a plurality of first tensors, wherein the first tensors comprise a plurality of feature tensors, and the feature tensors correspond to a plurality of features of a face;
a shared decoder module having a plurality of second parameters and configured to receive the feature tensors so as to generate a plurality of feature images;
a synthesis module configured to receive the first tensors so as to generate a vector; and
a classification module configured to receive the vector so as to generate a classification:
the at least one processing unit is configured to perform in a training epoch:
(a) repetitively executing: taking a training image from a training set as the input image: obtaining a first loss based on a plurality of training feature images of the training image and the feature images corresponding to the training image; and obtaining a second loss based on a classification label of the training image and the classification generated by the classification module in correspondence with the training image; and
(b) updating the first parameters and the second parameters based on an average value of all of the first losses obtained in the step (a), an average value of all of the second losses obtained in the step (a), and an updating algorithm.