Novel Learning-Based Visual Algorithms and Fusion Methods for High-Dimensional/Multi-Modality Big Data

基于学习的新型高维/多模态大数据视觉算法和融合方法

• 批准号: RGPIN-2022-02948
• 负责人: Wu, Jonathan
• 金额: $ 2.4万
• 依托单位: University of Windsor
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2022
• 资助国家: 加拿大
• 起止时间: 2022-01-01 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=752314
• 关键词: Novel; Learning; Based; Visual; Algorithms

During the last 10 years we have witnessed the immense growth of machine learning (ML) and deep learning (DL) due to the collection of vast amounts of labeled data and the ever-continuing innovation of high-performance computers (HPCs). ML and DL have dramatically become an integral part of our lives and are capable of performing tasks with precision like never before. They, in conjunction with computer vision, have become the cornerstone of a multitude of applications, including customs and border protection (CBP), intelligent transportation systems (ITS), video surveillance, vision-context applications, remote sensing and geographic information systems (GIS), biomedical image analysis, bioinformatics, and information security. The performance of many algorithms in real-world applications depends largely on a strong representation of their input streams. In most cases, unimodal feature representation will be biased and inadequate for a certain learning task, while the multi-modal feature learning frameworks can overcome the aforesaid shortcomings through learning complementary clues. However, the combined feature vector of multi-model systems often lies in a high- dimensional space, posing a serious problem in the final pattern recognition. At the same time, such high-dimensional feature vectors can include redundant components and noise. Thus, it is crucial to develop novel learning-based modeling technologies and fusion methods to remove redundancy and noise and handle complex big data. This proposal focuses on designing and developing hierarchical representation learning models and fusion methods for big data analysis, e.g., multi-modal medical records, static images, time/frequency-domain signals, and dynamic video streams. Specifically, the representation learning strategies will be developed based on machine/deep learning and other relevant models, e.g., non-iterative Moore-Penrose inverse (MPI) networks comprised of random vector functional links (RVFL) and Deep Convolutional Neural Networks (DCNNs). We believe that the research outcomes of the proposed themes will have a significant impact on the academic, healthcare, research, and adjacent industries in Canada and around the world as well in the related fields. The knowledge transfer of this proposal will be carried out through the dissemination of publications in international conferences, journals, books/book chapters, and the PI's active industrial collaborations. At the same time, the HQPs involved in this project will become leaders with cutting-edge expertise in academia as well as related industries through world-class training in the areas of machine learning, computer vision, and image/biomedical image processing and analysis under an interdisciplinary setup.

在过去的10年里，我们见证了机器学习（ML）和深度学习（DL）的巨大增长，这是由于大量标记数据的收集和高性能计算机（HPC）的不断创新。ML和DL已经戏剧性地成为我们生活中不可或缺的一部分，并且能够以前所未有的精度执行任务。它们与计算机视觉一起，已成为众多应用的基石，包括海关和边境保护（CBP），智能交通系统（ITS），视频监控，视觉环境应用，遥感和地理信息系统（GIS），生物医学图像分析，生物信息学和信息安全。许多算法在实际应用中的性能很大程度上取决于其输入流的强表示。在大多数情况下，单模态特征表示对于特定的学习任务来说是有偏差和不充分的，而多模态特征学习框架可以通过学习互补线索来克服上述缺点。然而，多模型系统的组合特征向量往往位于高维空间，给最终的模式识别带来严重的问题。同时，这种高维特征向量可能包括冗余分量和噪声。因此，开发新的基于学习的建模技术和融合方法来消除冗余和噪声并处理复杂的大数据至关重要。该提案的重点是设计和开发用于大数据分析的分层表示学习模型和融合方法，例如，多模态医疗记录、静态图像、时域/频域信号和动态视频流。具体来说，表示学习策略将基于机器/深度学习和其他相关模型来开发，例如，非迭代Moore-Penrose逆（MPI）网络由随机向量函数链接（RVFL）和深度卷积神经网络（DCNN）组成。我们相信，拟议主题的研究成果将对加拿大和世界各地的学术，医疗保健，研究和邻近行业以及相关领域产生重大影响。本提案的知识转移将通过在国际会议、期刊、书籍/书籍章节中传播出版物以及PI积极的行业合作来进行。与此同时，参与该项目的HQP将通过跨学科设置下的机器学习，计算机视觉和图像/生物医学图像处理和分析领域的世界一流培训，成为学术界和相关行业的前沿专业知识的领导者。