US 11,704,772 B2
Image classification system
Xiaoai Jiang, Lexington, MA (US); Theodor Ross, North Reading, MA (US); Suzanne Baker, Lincoln, MA (US); Matthew Campbell, Santa Monica, CA (US); and Cory Larsen, Cambridge, MA (US)
Assigned to Raytheon Company, Waltham, MA (US)
Filed by Raytheon Company, Waltham, MA (US)
Filed on Nov. 19, 2020, as Appl. No. 16/952,596.
Prior Publication US 2022/0156885 A1, May 19, 2022
Int. Cl. G06T 3/60 (2006.01); G01S 13/90 (2006.01); G06N 3/08 (2023.01); G06F 18/22 (2023.01); G06F 18/214 (2023.01); G06F 18/2411 (2023.01); G06N 3/045 (2023.01)
CPC G06T 3/60 (2013.01) [G01S 13/9064 (2019.05); G06F 18/214 (2023.01); G06F 18/22 (2023.01); G06F 18/2411 (2023.01); G06N 3/045 (2023.01); G06N 3/08 (2013.01)] 17 Claims
OG exemplary drawing
 
1. A method comprising:
obtaining an image;
identifying a rotation angle for the image by processing the image with a first neural network;
rotating the image by the identified rotation angle to generate a rotated image;
classifying the image with a second neural network; and
outputting an indication of an outcome of the classification,
wherein the first neural network is trained, at least in part, based on a categorical distance between training data and an output that is produced by the first neural network, the categorical distance being based on a distance between an index of a largest element in an output vector that is generated by the first neural network for a training image and an index corresponding to a label that is associated with the training image, and
wherein the categorical distance is defined as follows;
lθ=min(((ξy−ξŷ)%K),((ξŷ−ξy)%K)),
where ξy is the index of the largest element in the output vector that is generated by the first neural network for the training image, K is a count of classes that are associated with the first neural network, and ξŷ is the index corresponding to the label that is associated with the training image, the index corresponding to the label begin an index of a one-hot representation of the label.