US 11,670,072 B2
Systems and computer-implemented methods for identifying anomalies in an object and training methods therefor
Negin Sokhandan Asl, Montreal (CA)
Assigned to SERVICENOW CANADA INC., Montreal (CA)
Filed by ELEMENT AI INC., Montreal (CA)
Filed on Oct. 2, 2020, as Appl. No. 17/62,004.
Prior Publication US 2022/0108122 A1, Apr. 7, 2022
Int. Cl. G06V 10/46 (2022.01); G06N 3/02 (2006.01); G06F 18/214 (2023.01); G06F 18/2415 (2023.01)
CPC G06V 10/469 (2022.01) [G06F 18/2155 (2023.01); G06F 18/2415 (2023.01); G06N 3/02 (2013.01)] 16 Claims
OG exemplary drawing
 
1. A computer-implemented method for identifying anomalies in an object, comprising:
supplying, to an image encoder of a system, an input image of the object, the input image of the object containing zero or more anomalies;
generating, at the image encoder, an image model; and
applying the generated image model to an image decoder of the system, the image decoder forming a substitute non-anomalous image of the object, differences between the input image of the object and the substitute non-anomalous image of the object identifying zero or more areas of the input image of the object that contain the zero or more anomalies;
wherein the system implements a flow-based model;
wherein the system has been trained using (a) a set of augmented anomaly-free images of the object applied at the image encoder and (b) a reconstruction loss calculated based on a norm of differences between each augmented anomaly-free image of the object and a corresponding output image from the image decoder;
wherein:
the flow-based model forms a Gaussian model in which errors have a null mean and a predetermined standard deviation;
the system has been trained in unsupervised mode by supplying the set of augmented anomaly-free images of the object to the image encoder and by using the mean and the standard deviation of the flow-based model;
the system has been trained further by calculating a log-likelihood loss based on the mean and standard deviation of the flow-based model;
the log-likelihood loss is calculated in part based on a ratio of an output of a current layer of the flow-based model over an output of a previous layer of the flow-based model; and
the system has been trained further by calculating a regularization loss based on a ratio of the output of the previous layer of the flow-based model over the output of the current layer of the flow-based model.