报告主题：Neural network for inverse problem in imaging: from supervised learning to self-supervised learning
报 告 人：纪辉 副教授（新加坡国立大学数学系）
报告时间：2020年10月19日（周一） 10:00
参会方式：腾讯会议
会 议 ID：374 827 820
会议链接：
https://meeting.tencent.com/s/9loitB5hCj4z 主办部门：上海大学运筹与优化开放实验室-国际科研合作平台、上海市运筹学会、上海大学理学院数学系
报告摘要：In last few years, deep learning has emerged as one powerful tool for solving many inverse problems in imaging. Most existing works are based on supervised learning which calls an external image dataset to train a neural network, and the trained model from these methods only work well on the test data whose noise level is the same as training data. In the first part of the talk, we will introduce a noise-blind deep neural network for image deconvolution which allows to train a universal model for processing data with unknown or varying noise level. The proposed network is built on the unrolling of variational expectation maximization method. In the second part, we will show that, even without any training data, a deep network still can learn how to solve inverse problems. In the context of image reconstruction of compressed sensing (CS), we will introduce a self-supervised deep network method for CS image reconstruction with state-of-the-art performance, which is built on the Bayesian approximation to MMSE estimator via Bayesian deep network.
报告人简介: Dr. Ji Hui is currently an Associate Professor in the department of Mathematics at NUS. He is also the director of Centre for Wavelets, Approximation and Information Processing (CWAIP) and the affiliated faculty member of Institute of Data Sciences. He received Ph.D. in Computer Science from the University of Maryland at College Park in 2006. His research interest covers computational harmonic analysis, computational vision, machine learning, and data sciences.
欢迎教师、学生参加！