JSAI2023

Presentation information

Poster Session

General Session » Poster session

[3Xin4] Poster session 1

Thu. Jun 8, 2023 1:30 PM - 3:10 PM Room X (Exhibition hall B)

[3Xin4-34] Gated Denoising Diffusion Probabilistic Model with Diffusion Process in Bounding Boxes

〇Kohei Makino1 (1.Toyota Technological Institute)

Keywords:Denoising Diffusion Probabilistic Model, Inpainting, Gate mechanism

An inpainting task, which interpolates a part of an image, is used to revise pictures. However, the model for inpainting is only available if the revision area is clear. As an initial study to address image revision for the unknown error, this study proposes a gated denoising diffusion probabilistic model (GDDPM) training with a bounded diffusion process, assuming a diffusion process in a bounding box on the image. GDDPM is a DDPM with a gating mechanism to detect modification parts and denoise noise inside it. In training, the gate classifies noisy areas in a bounded diffusion process. Our experiments on CIFAR-10 show that GDDPM with gate training is necessary gate training.

Authentication for paper PDF access

A password is required to view paper PDFs. If you are a registered participant, please log on the site from Participant Log In.
You could view the PDF with entering the PDF viewing password bellow.

Password