US 11,657,274 B2
Weakly-supervised semantic segmentation with self-guidance
Kunpeng Li, Somerville, MA (US); Ziyan Wu, Princeton, NJ (US); Kuan-Chuan Peng, Plainsboro, NJ (US); and Jan Ernst, Princeton, NJ (US)
Assigned to SIEMENS AKTIENGESELLSCHAFT, Munich (DE)
Appl. No. 16/760,096
Filed by Siemens Aktiengesellschaft, Munich (DE)
PCT Filed Oct. 9, 2018, PCT No. PCT/US2018/054993
§ 371(c)(1), (2) Date Apr. 29, 2020,
PCT Pub. No. WO2019/089192, PCT Pub. Date May 9, 2019.
Claims priority of provisional application 62/581,321, filed on Nov. 3, 2017.
Prior Publication US 2020/0356854 A1, Nov. 12, 2020
Int. Cl. G06N 3/08 (2023.01); G06N 5/04 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 5/04 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method, comprising:
providing an input image to a convolutional neural network (CNN), wherein the input image is associated with an image-level label;
obtaining, from the CNN, a first classification probability distribution associated with the input image;
determining, based at least in part on the first classification probability distribution and the image-level label of the input image, a classification loss;
determining an attention map associated with the input image;
applying a thresholding operation to the attention map to obtain a soft mask;
applying the soft mask to the input image to obtain a masked image;
providing the masked image as an input to the CNN;
obtaining, from the CNN, a second classification probability distribution associated with the masked image;
determining an attention mining loss associated with the attention map based at least in part on the first classification probability distribution and the second classification probability distribution; and
utilizing the classification loss and the attention mining loss to self-guide training of the CNN.