主讲题目：A Journey to Single Image Defocus Deblurring: From Two-Stage Deconvolution to  End2End Kernel Mixture Learning

主讲人：全宇晖

时间：2024年5月8日上午10:30

腾讯会议号：446-864-006

报告简介： 

Single image defocus deblurring (SIDD) refers to recovering an all-in-focus image from a defocused blurry one. This task is particularly challenging due to the spatially-varying nature of defocus blur effects, which can vary significantly in size across the image. This talk provides an overview of our studies on SIDD. It begins with the introduction on two-stage SIDD, involving a defocus map estimation step followed by a non-blind deconvolution step. Our work on both steps will be presented, including a patch rank-based defocus map estimator and two unsupervised deep nonblind deconvolution methods. The latter includes a zero-shot learning method that eliminates the need for training samples and a fully-unsupervised learning method that removes the requirement for ground-truth clear images. Subsequently, we will delve into our recent work on kernel mixture model-based end-to-end learning for SSID. Four studies along this line will be presented. The first three focus on defocus blur modelling via a kernel mixture model and design network architecture through efficient deep unrolling techniques. The last study shows that inverse kernels be can directly and effectively modelled using a mixture of implicit neural kernels, based on which a more efficient neural network for SIDD is designed. Finally, potential future directions of SIDD will be discussed.

主讲人简介：

全宇晖：华南理工大学计算机学院副教授、博士生导师，入选2017年广东省珠江人才计划青年拔尖人才计划，OPPO计算影像专家联谊会委员，主要研究方向为图像处理、深度学习、稀疏表达，已发表高水平论文近80篇（包括CVPR、ICCV、NeurIPS、AAAI、ECCV等计算机视觉与机器学习国际顶级会议论文，以及TPAMI、IJCV、TIP、TSP、PR等人工智能与图像处理顶国际级期刊论文），主持国家自然科学基金项目三项、广东省自然科学基金项目三项、广州市科技计划项目一项，参与十多项国家科研项目，获2019 年广东省科学技术进步二等奖（第二完成人）、2017年度广东计算机学会青年学术秀一等奖、2016年度ACM广州分会学术新星奖、2016年度广东省计算机学会优秀论文一等奖、2014年度ACM 广州分会优秀博士论文奖。（个人主页：csyhquan.github.io）