US 12,223,764 B2
System and method for rotation invariant fingerprint recognition
Zhinong Li, Alpharetta, GA (US); and Xiaowu Zhang, Alpharetta, GA (US)
Assigned to ARMATURA LLC, Alpharetta, GA (US)
Filed by Armatura LLC, Alpharetta, GA (US)
Filed on Feb. 28, 2022, as Appl. No. 17/682,923.
Prior Publication US 2023/0274574 A1, Aug. 31, 2023
Int. Cl. G06V 40/12 (2022.01); G06F 21/32 (2013.01); G06V 10/44 (2022.01); G06V 10/82 (2022.01)
CPC G06V 40/1347 (2022.01) [G06F 21/32 (2013.01); G06V 10/454 (2022.01); G06V 10/82 (2022.01)] 18 Claims
OG exemplary drawing
 
1. A method implemented on at least one processor, a memory, and a communication platform enabling connection to a network for biometric authentication, comprising:
generating rotation invariant (RI) training data with rotation invariant characteristics based on a plurality of training samples, each of which includes an image capturing biometric information of a person, a feature vector extracted from the image, and a label indicative of an identity of the person;
conducting machine learning based on the RI training data to learn weights of each of rotation covariant (RC) convolution kernels for each of a plurality of convolution layers;
receiving an input image with fingerprint information captured therein related to a person to be authenticated;
preprocessing the input image to obtain an initial feature map;
at each of the plurality of convolution layers, processing a feature map based on the RC convolution kernels for the layer to output an RC feature map, wherein the feature map is one of the initial feature map and an output RC feature map from a previous one of the plurality of convolution layers, and the RC feature map output by the last one of the plurality of convolution layers corresponds to a feature vector representing fingerprint features of the person in a rotation invariant manner; and
authenticating the person based on the feature vector.