IEEE出版7天征稿倒计时-连续2届快至会后4个月EI检索-第三届图像处理、多媒体技术与机器学习国际学术会议（IPMML 2026）

特邀报告：
Prof. Ling Guan

Toronto Metropolitan University, Canada
IEEE Life Fellow| CAE Fellow| EIC Fellow| IEEE Canada C.C. Gotlieb Medel
Ling Guan教授是国际知名的多媒体信号处理和机器学习领域学者，曾任加拿大一级研究讲座教授及多所大学教授，并创立多媒体研究实验室。他与Kuo教授共同发起“超越深度学习”的新研究方向。林博士担任多项IEEE期刊编委，主导创办了IEEE泛太平洋多媒体会议，并获得IEEE会士、加拿大工程院院士等荣誉，以及IEEE最佳论文奖和C.C.戈特利布奖章。

Speech Title: Learning beyond Deep Learning: Mathematics-Inspired Models for Multi-View Analysis

Abstract:

Acknowledging the tremendous contributions deep learning (DL) made to a broad range of image, video and multimedia processing tasks, further optimizing quality of the DL-based features has consistently presented a challenge, especially when working with multi-view data, arguably the most natural data format. Though DL models have been able to extract high quality features, jointly processing these features is largely confined to late fusion (score level and decision level) when information has been condensed into a few numbers, thus limiting quality of performance. In this talk, we present recent development from a significantly different perspective, a mathematics-inspired model; an approach belonging to the cooperative learning paradigm. Particularly, we call upon a natural multi-view processing architecture, discriminant correlation analysis, setting the stage for the development of an innovative platform. Due to its power to handle information with multiple views, the platform is termed as discriminant multiple correlation (DMC) analysis. Depending on the nature of the data sources to work with, DMC features two distinct designs, one with a perceptron-style NN (PNN) to handle the nonlinear processing part, and the other with a convolution-style NN (CNN). Statistics collected from experiments of numerous multi-view analysis and recognition benchmarks evidently show that the MI models generate impressive (sometimes unprecedented) performance accuracies.

Prof. Hideo Saito

Keio University, Japan

IEICE Fellow| VRSJ Fellow| President of ITE| Former President of IEICE ISS

Hideo Saito 博士是庆应义塾大学正教授，3D 计算机视觉和 AR/MR 领域知名学者，曾参与卡内基梅隆大学“虚拟化现实”项目。现任该校多个学术领导职务及 IEICE 和 ITE 等重要学会会长，并多次担任国际顶级会议主席，获得 ISMAR 2023 最佳期刊论文奖等荣誉。

Keynote 1: Journey to the Next Era: AI-driven Multi-view Analysis and the Future of Spatial Intelligence

Abstract：

Since 1997, our exploration has been fueled by the ambition to reconstruct 3D worlds from large-scale multi-view camera arrays, giving birth to “free-viewpoint video.” For over two decades, we have refined the foundational pillars of multi-view geometry, such as camera calibration and SLAM. However, we are now entering a transformative “Next Era” where the synergy between these traditional methodologies and modern AI is redefining the boundaries of what is possible.

In this talk, I will present our journey as it evolves from early 50-camera setups to the cutting edge of AI-driven 3D vision. I will discuss how recent breakthroughs in neural rendering and generative video intelligence are unleashing the true potential of multi-view systems—moving beyond mere geometric reconstruction toward deep "Spatial Intelligence."

While these technologies have found impactful applications in fields like surgical scene analysis and Medical AR, their potential extends far beyond, promising a future where AI can perceive, understand, and interact with complex 3D environments in real-time. I will introduce our latest research projects and share my excitement for this new chapter in computer vision, where multi-view analysis serves as the cornerstone for the next generation of intelligent systems.

Prof. Mohan Kankanhalli

National University of Singapore, Singapore

ACM FellowIEEE Fellow | IAPR FellowSNAS Fellow | Deputy Executive Chairman of AI Singapore | Director of NUS AI Institute
Mohan Kankanhalli 教授，新加坡国立大学（NUS）人工智能研究院院长，现任教务长讲席计算机科学教授。他同时担任新加坡国家人工智能研发项目“AI Singapore”的人才培养副执行主席。2016 至 2022 年间，他曾任新加坡国立大学计算机学院院长。在此之前，他于 2014 至 2016 年担任 NUS 研究生教育副校长，2011 至 2013 年任协理副校长；更早前，他在 2008 至 2010 年任计算机学院学术事务与研究生教育副院长，2001 至 2007 年任研究副院长。

Kankanhalli 教授的研究兴趣包括多模态计算、计算机视觉和可信人工智能。他在图像与视频理解、视觉显著性和可信 AI 等领域做出了基础性贡献。现任《IEEE 多媒体》杂志副主编，并担任《Springer 多媒体系统》和《Springer 大数据杂志》编委。

Kankanhalli 教授曾担任世界经济论坛 2023–2024 年度全球人工智能未来理事会成员，现为 ACM 全球技术政策理事会成员。他是 IEEE、IAPR 和 ACM 的会士。

Keynote 2: TBD
Abstract：TBD

Prof. Tülay Adalı

University of Maryland, Baltimore County (UMBC), USA
IEEE Fellow | AIMBE Fellow | AAIA Fellow | Fulbright Scholar | EiC of IEEE SPM

美国UMBC杰出大学教授，机器学习与信号处理实验室主任。研究方向：统计信号处理、机器学习、医学图像分析。现任《IEEE信号处理杂志》主编，曾任IEEE SPS技术方向副主席，IEEE、AIMBE、AAIA会士，洪堡研究奖得主，IEEE SPS最佳论文奖。

Keynote 3: Stable Components in High-Dimensional Data: Joint Decomposition Across Datasets

Abstract
In many fields today, such as neuroscience, remote sensing, computational social science, and the physical sciences, multiple datasets are readily available. Matrix and tensor factorizations provide a principled framework for joint analysis, enabling the fusion of such datasets so that they can interact and inform each other while making relatively modest assumptions about their underlying relationships. A central motivation for these methods is the recovery of components that are directly interpretable and, in favorable settings, comparable across datasets, subjects, or measurement sessions. This naturally raises questions of reproducibility and stability of the estimated components, particularly in relation to more flexible representation learning approaches.
This talk presents an overview of matrix and tensor factorization models, with emphasis on independent component analysis and its multi-dataset extension, independent vector analysis. We discuss how these approaches identify a directly interpretable latent structure, and how resulting components can be analyzed across modalities and observations. Examples from multi-subject neuroimaging data illustrate how recurring components can be identified across individuals and used in applications such as subgroup discovery for precision medicine. We also briefly contrast these properties with those of modern deep learning methods in terms of interpretability and identifiability, and highlight remaining challenges related to robustness, model selection, and practical reproducibility.