公开(公告)号：US11003959B1

公开(公告)日：2021-05-11

申请号：US16440384

申请日：2019-06-13

Applicant: AMAZON TECHNOLOGIES, INC.

Inventor： Ilya Levner ,  Konstantinos Boulis ,  Gurbinder Gill ,  Canku Calargun ,  Prajwal Yadapadithaya ,  Venkata Krishnan Ramamoorthy ,  Zhaoqing Ma

IPC: G06K9/62 ,  G06N3/04 ,  G06N3/08 ,  G06K9/68

Abstract: Categorizing images may include training a first neural network to cluster a plurality of images to obtain a first image embedding space, wherein a vector representation is determined for each of the plurality of images based on the training, determining a vector norm value corresponding to each of the plurality of images based on the vector representation for each of the plurality of images, and identifying a first subset of the images for which a corresponding vector norm value satisfies a predetermined vector norm quality threshold. Then, a second neural network may be trained using the first subset of images to obtain a second image embedding space, and the second image embedding space may be used to categorize additional images.