US 11,989,972 B2
Method for predicting characteristic information of target to be recognized, method for training neural network predicting characteristic information of target to be recognized, and computer-readable storage medium storing instructions to perform neural network training method
Hyogi Lee, Seongnam-si (KR); and Kideok Lee, Seongnam-si (KR)
Assigned to Suprema AI Inc., Seongnam-si (KR)
Filed by Suprema AI Inc., Seongnam-si (KR)
Filed on Mar. 9, 2023, as Appl. No. 18/181,322.
Claims priority of application No. 10-2022-0125645 (KR), filed on Sep. 30, 2022.
Prior Publication US 2023/0215216 A1, Jul. 6, 2023
Int. Cl. G06V 40/16 (2022.01); G06V 10/77 (2022.01); G06V 10/774 (2022.01); G06V 10/82 (2022.01)
CPC G06V 40/171 (2022.01) [G06V 10/7715 (2022.01); G06V 10/774 (2022.01); G06V 10/82 (2022.01); G06V 40/178 (2022.01)] 8 Claims
OG exemplary drawing
 
1. A method for predicting characteristic information of a target person to be recognized, the method comprising:
acquiring a face image; and
predicting the characteristic information of the target person to be recognized from the face image using a first pre-trained neural network,
wherein the face image comprises a face image of the target person wearing a mask,
wherein the characteristic information includes gender, age, or race of the target person contained in the face image,
wherein the first pre-trained neural network is trained based on a plurality of first face images, a plurality of second face images, and characteristic information on each of the plurality of first face images, and on each of the plurality of second face images,
wherein the plurality of first face images are images acquired by capturing non-covered faces, and
wherein the plurality of second face images are images synthesized, by a second pre-trained neural network, by synthesizing each of a plurality of mask images with each of the plurality of first face images, wherein each of the plurality of mask images represents a distinctive type, color and shape of mask, wherein the second pre-trained neural network is trained to synthesize a mask image with a face image, and wherein the synthesizing each of the plurality of mask images with each of the plurality of first face images comprises extracting a plurality of facial feature points from the each of the plurality of first face images, determining a plurality of mask feature points for each of the plurality of mask images that represent a set of salient geometry points of the each of the plurality of mask images, matching the plurality of mask feature points in a plurality of sizes of the each of the plurality of mask images with the plurality of facial feature points in a plurality of locations of the each of the plurality of first face images to produce the plurality of second face images.