公开(公告)号：US12182918B2

公开(公告)日：2024-12-31

申请号：US17847472

申请日：2022-06-23

Applicant: KOREA ADVANCED INSTITUTE OF SCIENCE AND TECHNOLOGY

Inventor： Junyong Noh ,  Ha Young Chang ,  Kwanggyoon Seo ,  Jung Eun Yoo

IPC: G06T13/20 ,  G06T7/194

Abstract: A method of training a deep neural network for generating a cinemagraph is disclosed. The method may include preparing a foreground layer input by using an input video, preparing a background layer input by using the input video, providing a foreground layer output and a background layer output from the DNN by inputting, to the DNN, the foreground layer input and the background layer input, providing an output video by synthesizing the foreground layer output with the background layer output, and updating intrinsic parameters of the DNN a plurality of times, based on the input video, the foreground layer input, the foreground layer output, the background layer output, and the output video.