US 12,272,046 B2
Feature detection based on neural networks
JingChang Huang, Shanghai (CN); Guo Qiang Hu, Shanghai (CN); Peng Ji, Nanjing (CN); Yuan Yuan Ding, Shanghai (CN); Sheng Nan Zhu, Shanghai (CN); and Jinfeng Li, Shanghai (CN)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Sep. 21, 2020, as Appl. No. 17/026,620.
Prior Publication US 2022/0092756 A1, Mar. 24, 2022
Int. Cl. G06N 3/08 (2023.01); G06N 3/045 (2023.01); G06T 3/40 (2006.01); G06T 7/00 (2017.01); G06V 20/40 (2022.01)
CPC G06T 7/001 (2013.01) [G06N 3/045 (2023.01); G06N 3/08 (2013.01); G06T 3/40 (2013.01); G06V 20/41 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A method, comprising:
inputting, into a set of neural networks, a plurality of different images of a same region of interest in a same object, wherein a visual feature is present in the region and detectability of the feature within each image of the region varies depending on a value of a variable condition under which that image is captured, wherein each image of the region has been captured under a different value of the variable condition, wherein the variable condition comprises at least lighting across a spectrum and two camera spherical coordinate viewing angles from which the plurality of different images are captured;
generating, by the set of neural networks, a classification for each image, wherein each classification includes a confidence score in a prediction of whether the visual feature is present in the region;
ensembling the classification for each image to generate a final classification for the region;
computing, by applying a loss function, a loss based on comparing the final classification to a ground truth of whether the feature is present in the region; and
adjusting parameters of the set of neural networks based on the computed loss.