US 12,249,120 B2
Method, system and storage media for training a graphics processing neural network with a patch-based approach
Chi-Chung Chen, Taipei (TW)
Assigned to AETHERAI IP HOLDING LLC, Frisco, TX (US)
Appl. No. 17/769,324
Filed by AETHERAI IP HOLDING LLC, Frisco, TX (US)
PCT Filed Aug. 26, 2020, PCT No. PCT/US2020/047856
§ 371(c)(1), (2) Date Apr. 14, 2022,
PCT Pub. No. WO2022/046041, PCT Pub. Date Mar. 3, 2022.
Prior Publication US 2024/0135677 A1, Apr. 25, 2024
Int. Cl. G06V 10/764 (2022.01); G06V 10/26 (2022.01); G06V 10/82 (2022.01)
CPC G06V 10/764 (2022.01) [G06V 10/267 (2022.01); G06V 10/82 (2022.01)] 29 Claims
OG exemplary drawing
 
1. A method for training a graphic processing neural network with a patch-based approach, the method performed by a system including at least a processor with embedded memory, the method comprising:
(a) one of the at least one processor calculating an overlapping size and an invalid size of an output of each of at least one of multiple feature extraction layers of the graphic processing neural network according to a predetermined cropping scheme;
(b) one of the at least one processor dividing an input image into multiple first patches in a patch pattern with each first patch having an augmented marginal portion overlapping other first patches adjacent to the first patch with the overlapping size;
(c) the at least one processor propagating the multiple first patches through the multiple feature extraction layers of the graphic processing neural network in a forward pass for each first patch to generate multiple feature maps, and cropping out an invalid portion from each of at least one feature map for each first patch according to the invalid size of the feature map and the predetermined cropping scheme; and
(d) one of the at least one processor aggregating the last feature maps associated with the first patches according to the patch pattern to generate a global embedding map and passing the global embedding map to classification layers of the graphic processing neural network in the forward pass for model prediction.