时间：2024/12/26 9:00-12:00

　　地点：数学与统计楼213

报告人：陈柯 教授      英国思克莱德大学

题目：Semi-supervised and self-supervised learning techniques for imaging applications 

摘要：Variational modelling has been effective for solving a large class of imaging tasks, 

as evidenced by many major papers in the exciting field. In this talk, I review two related works that use deep learning in order to deliver a fast implementation, when the number of trained data is large enough. The central idea in this talk is how to deal with small or no training data using the data approach.

Firstly, for image segmentation, accuracy is correlated with size of trained data. In digital pathology, there exist problems where it is either time-consuming to acquire enough trained data or impossible to get any at all. By semi-supervision, we present our preliminary work towards using variational models to assist the augmentation of training data and hence to increase the accuracy of learning methods or to enable effective learning.
We also discuss more recent attempts on generative models.

Secondly, for image denoising, it is possible to learn effective restoration without any training data by self-supervision. Based on this idea, we developed a dual domain learning method for reconstruction of denoised images using low dose (and noisy) CT sinograms as input. Tests can show that self-supervised results are almost as good as fully supervised learning. 

报告人简介：陈柯，教授，英国思克莱德大学数学与统计系主任，大连理工大学百人计划专家，英国IMA Fellow、御批数学家。陈柯教授作为CMIT和LCMH创新团队负责人，主要研究方向为计算数学、应用数学、图像处理、医疗应用等，近年来主要研究图像处理中各种反问题的模型构造，算法和应用。在SIAM Journal on Imaging Sciences、IEEE TIP、JCP等国际权威期刊上发表超过170余篇学术论文。目前担任Numerical Algorithms等多个国际知名期刊的编辑及执行编辑。