Logo

How Neural Network can Learn to Denoise an Image without any Training Data 151

202e0975dc2efc0510668f1213396dcc4febfa28

Speaker

Hui Ji, National University of Singapore, Singapore

Time

2020.07.08 14:00-15:00

Venue

Online—ZOOM APP

ZOOM Info

ZOOM Link

Conference ID: 924-778-64329

PIN Code: 414615

Abstract

In last few years, deep learning has emerged as one powerful tool for image denoising, a fundamental problem in image processing. Most existing works are based on supervised learning which calls an external image dataset to train a denoising network. In many real scenarios, the instruction of an unbiased and comprehensive image dataset can be a challenging and sometimes impossible task, e.g., medical and scientific imaging. Contradict to popular belief, we will show in this talk that, without seeing any other image data, a deep network still can learn how to denoise an image with solid performance. In this talk, we will introduce a self-supervised image denoising method with state-of-the-art performance, which is built on two concepts: (1) Bernoulli-sampling-based data augmentation and (2) dropout-based training/testing scheme.

It is a joint work with Quan Yuhui, Chen Mingqin from SCUT, China and Pang Tongyao from NUS.

Video

How Neural Network can Learn to Denoise an Image without any Training Data

Slides

Slides

Sponsors