公开(公告)号：US11961277B2

公开(公告)日：2024-04-16

申请号：US17562980

申请日：2021-12-27

Applicant: Tsinghua University

Inventor： Gao Huang ,  Shiji Song ,  Haojun Jiang ,  Le Yang ,  Yiming Chen

IPC: G06V10/00 ,  G06F18/22 ,  G06V10/44 ,  G06V10/77 ,  G06V10/774

CPC classification number: G06V10/443 ,  G06F18/22 ,  G06V10/7715 ,  G06V10/774

Abstract: A method for detecting image information includes: acquiring at least one sample of image pair to be processed; calculating a reconstruction loss function of the second feature extraction model based on the first image samples and the first reconstructed image feature information; calculating an adversarial loss function of the third feature extraction model based on the second reconstructed image feature information and the first image samples; optimizing the first model parameters in the first feature extraction model based on the reconstruction and the adversarial loss function to generate the optimized first feature extraction model; inputting the acquired image pair to be processed into the optimized first feature extraction model to generate the difference information. The method reduces the first feature extraction model's dependence on the labeled data and improves the model's recognition efficiency and accuracy by using the samples without the labeled difference information.