会议日程

王卫卫个人简介

西安电子科技大学数学与统计学院教授，博导。主要研究方向：机器学习、图像处理的变分偏微分方程方法。主要成果：合作出版科研专著《图像处理的变分与偏微分方程方法》一部，在国内外重要学术期刊与会议上合作发表论文60余篇，SCI检索论文20余篇，发表期刊包括IEEE Trans. on Image Processing，IEEE Trans. On Circuits and Systems for Video Technology, SIAM J. on Multiscale Modeling and Simulation，Pattern Recognition,Neurocomputing，SignalProcessing等；主持完成国家自然科学基金面上项目2项；获得陕西省高校科技奖2项，陕西省科技奖1项。

图像恢复和边缘提取的联合算法

报告题目：图像恢复与边缘提取的联合合作博弈。

摘要：图像恢复和边缘提取是图像处理的两大基本问题，在退化图像中提取边缘颇具挑战性，联合实现图像恢复和边缘提取很有意义。针对图像恢复，利用已有高效的图像去噪方法定义了一个新的启发式图像正则项；针对边缘提取，改进了一个求图像梯度的优化模型；将两者结合，给出一个合作博弈模型来联合实现图像恢复和边缘提取；给出一个求解模型的迭代算法，证明了其收敛性；在模糊图像和缺失像素图像上的实验结果表明，所提模型和算法在图像恢复和边缘提取两方面都具有良好的性能。

刘君个人简介

北京师范大学副教授，博士生导师。主要研究方向为变分法及深度学习相关的图像处理算法与应用。目前发表论文20余篇，相关研究结果在图像处理与计算机视觉领域权威期刊如IJCV、IEEE TIP、SIIMS等发表。研究成果曾获教育部高等学校优秀科研成果二等奖（第三完成人），毕业论文曾获全国优秀博士毕业论文提名(2013年度)。主持参与多项国家科研项目。

报告题目：Deep Learning based Image Segmentation and Restoration with Prior of Variational Model

摘要：Deep Convolutional Neural Networks (DCNN) have achieved prominent performance in a series of imageprocessing problems.However, in image segmentation and restoration, DCNN predict the class of each pixel independently tasks. The spatial regularity of the objects in the image is still a problem for these methods. Especially whengiven few training data, DCNN could not perform well in the details. The isolated and scattered small regionsoften appear in all kinds of DCNN segmentation results. The texture restoration is not good enough since lack of nonlocal information in the image restoration. In this talk, we propose a method to consider different priorsincluding regularization and volume persevering in the deep learning method. To be different frompostprocessing technique in many existing methods,we design some new layers for DCNN using the idea of prime-dual method in the variational method. As applications,we apply these proposed layers to several popularsegmentation and restoration DCNNs such Deeplab3+, Unet, DnCNN, the experimental results show that our method can achieve state-of-the-art results on some public datasets such as pascal voc2012.

殷钶个人简介

2013年获美国佐治亚理工学院应用数学专业博士，之后在美国加州大学洛杉矶分校数学系从事博士后研究工作。 2016年加入华中科技大学数学中心，任副研究员。研究方向为基于优化和偏微分方程的反问题解法，及其在图像重构中的应用。近几年开始研究优化和变分原理在计算材料科学以及机器学习中的应用，相关论文已被Communications in Mathematical Sciences， Journal of scientific computing 等收录。

报告题目：A new smooth approximation of the maximum function and its applications to imaging problems

摘要：In this talk, we present and analyze a smoothing technique  for the maximum function, based on the compensated convex transforms, originally proposed by Kewei Zhang. This type of approximation techniques has an analytical formula which is a $C^{1,1}$ function and a tighter approximation than the log-sum-exp approximation. Moreover, it preserves convexity. Many inverse problems and optimization problems can be formulated as minimization of the maximum of several (convex) functions. We propose smooth approximations to these (non-smooth) minimax problems, which allow first order optimization techniques, and the approximation error can be conveniently estimated. We demonstrate the effectiveness of the proposed method through numerical examples in Pott's model for multiphase partition, robust optimization, the computation of the Euclidean distance between two closed convex sets, and in implicit shape blending.

庞志峰个人简介

河南大学数学与统计学院副教授，硕士生导师。目前任河南省数字图形图像学会常务理事和常务副秘书长，并分别兼任该学会的智能精准放疗专业委员会和智能信息融合专业委员会副主委，同时任中国生物医学工程学会医学人工智能专委会青年委员会委员。主要研究图像处理中的数学理论与数值算法。主持国家自然科学基金1项，参与国家自然科学基金2项，国家重点基础研究发展计划(973项目)1项，授权专利1项。

报告题目：Image decomposition and restoration based on the adaptive direction total variation regularization

摘要：With the development of the deep learning, we need many labeled data to improve the learning  accuracy by employing some efficient numerical methods without any manual work. So the data clustering method has received extensive attention in these fields. In this report, we want to propose a new clustering method via modifying the classic Cheeger cut model. Since the model is nonconvex and nonsmooth, we use the linearization scheme to transform it into the easily solvable optimization and then use the ADMM to solve it. Numerical experiments demonstrate that the proposed method is competitive with the current state of the art in some data clustering methods.

陈柯个人简介

英国利物浦大学终身教授，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，在SIA M Journal on Imaging Sciences、IEEE Transactions on Image Processing、Journal of Computational Physics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，Journal of Mathematical Research and Exposition，Journal of Applied Mathematics ，International Journal of Computer Mathematics，Journal of Mathematical Research with Applications等多个国际知名期刊的编辑及执行编辑。