项目名称: 基于稀疏非一致多核学习的低分辨率视频识别研究
项目编号: No.61203248
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 自动化学科
项目作者: 任传贤
作者单位: 中山大学
项目金额: 25万元
中文摘要: 以图像与视频为主要媒介的生物特征识别是模式识别与自动化应用领域的前沿研究方向,也是人类在基础理论与应用研究中面临的重要挑战之一。 本项目以稀疏表示与多核学习方法为数学工具,以真实远距离监控视频和多类别数据为主要实验对象,认真探讨低分辨率数据的视觉不变特征生成原理,致力于研究具有群组效应的稀疏正则化方法、非一致核函数的局部保持匹配性能和大规模快速优化算法,提出实际有用的参数选择方法,在非一致多核学习、目标求解和在线快速算法方面具有较大创新性。 项目详细讨论Lpq范数对噪音和例外点的鲁棒性、特征选择的稀疏性和目标函数的光滑性。重点解决惩罚项导致的目标函数优化问题、多模态数据的相似度量学习与特征融合问题和大规模数据面临的快速计算问题,提高系统的判别特征提取能力和泛化能力。 通过本项目的研究,进一步加强稀疏多核学习框架的理论深度和应用广度,促进新方法向工业化与市场化方向转化。
中文关键词: 多尺度分析;核方法;判别分析;稀疏正则化;人像识别
英文摘要: Biological feature recognition based on the image and video media is the forefront research direction in the pattern recognition and automation application communities, and it is one of the most important challenges for the basic theories and applications that human facing. The program will exploit sparse representation and multi-kernel learning methods as the mathematical tools, focus on the real far-distant surveillance camera video and the multi-subject data, probe deeply into the principles of vision-invariant feature generation, and study the sparsity regularization with group effective, locality preserving capabilities of inconsistent kernel functions and the fast computation algorithm with large scale. The program will propose efficient and effective approaches for parameters selection and model evaluation, and present some novelties in inconsistent multi-kernel learning, objective solution and online computation. The robustness for noise and outliers, sparsity for feature selection and the smoothness for objective function due to Lpq-norm will be analyzed, then the pivotal problems including the objective optimization, similarity metric learning and feature fusion of multi-modal data, and the fast computation in large scale would be highlighted for improving the capabilities of feature extraction and g
英文关键词: Multi-scale analysis;Kernel method;Discriminant analysis;Sparsity regularization;Face recognition