US 12,272,133 B2
Automatic method to determine the authenticity of a product
Paolo Pegurri, Milan (IT); Marco Romelli, Milan (IT); and Luca Losa, Milan (IT)
Assigned to EBAY INC., San Jose, CA (US)
Filed by eBay Inc., San Jose, CA (US)
Filed on Aug. 12, 2021, as Appl. No. 17/444,932.
Claims priority of application No. 102020000020218 (IT), filed on Aug. 17, 2020.
Prior Publication US 2022/0051040 A1, Feb. 17, 2022
Int. Cl. G06V 20/80 (2022.01); G06N 3/08 (2023.01); G06T 7/00 (2017.01); G06V 10/20 (2022.01); G06V 10/94 (2022.01); G06V 20/00 (2022.01)
CPC G06V 10/95 (2022.01) [G06N 3/08 (2013.01); G06T 7/0004 (2013.01); G06V 10/255 (2022.01); G06V 20/80 (2022.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30112 (2013.01); G06T 2207/30124 (2013.01); G06V 20/95 (2022.01)] 17 Claims
OG exemplary drawing
 
1. An automatic method to determine authenticity of a product (P), the method comprising:
preparing a training dataset, the preparing comprising:
accessing a plurality of source images (IA0(m)) corresponding to an authentic product and a plurality of source images (IF0(n)) corresponding to a fake product,
annotating an area comprising a distinguishing feature (DPF) of the authentic product or the fake product in each of the source images (IA0(m), IF0(n)), and
enriching the training dataset by modifying each of the source images (IA0(m), IF0(n)) using a predefined augmentation algorithm to generate a plurality of training images (IA(m′), IF(n′)), the modifying comprising the predefined augmentation algorithm making small random changes to one or more attributes of the distinguishing feature (DPF) of each of the source images, the small random changes including one or more non-linear distortions that change an original shape of an object associated with the distinguishing feature (DFP) within at least some of the source images (IA0(m), IF0(n));
training at least one neural network with the training dataset;
accessing an input image (IU(k)) representative of the product (P) of a user to be analyzed; and
querying the at least one neural network (N) previously trained so that the latter is independently able to assign, for each attribute and/or for each distinguishing feature (DPF) of the input image (IU(k)), an authenticity index (R) representative of a probability of authenticity of the analyzed product (P).