项目名称: 深度属性特征学习及其应用研究
项目编号: No.61473256
项目类型: 面上项目
立项/批准年度: 2015
项目学科: 自动化技术、计算机技术
项目作者: 王东辉
作者单位: 浙江大学
项目金额: 80万元
中文摘要: 深度属性特征学习方法是一个崭新的研究方向,其主要目标是研究从低层多模态特征到高层属性特征的统一的深度学习架构,提出创新性的方法、模型和求解策略,并在具有抽象语义属性的艺术品检索问题上实现探索性示范应用。具体研究内容包括分析和评估单模态深度学习方法的性能特点,研究优化的单模态深度学习模型,提出具有更高语义表达的单模态特征深度学习方法;从多模态特征融合的角度,研究有效的深度学习架构和求解策略,为进一步的深度属性特征学习打下基础;通过对属性特征的层次分解,构建多模态特征到属性特征映射的深度学习架构及求解方法,实现更有效的属性预测和语义注解;通过在典型属性数据集和特定属性数据集上的应用研究,评估提出方法在典型应用任务中的具体性能指标。深度属性特征学习方法的研究成果不仅对机器学习新方法的研究具有重要的促进作用,而且能够为大规模特征学习、语义标注与内容理解、跨模态语义检索等诸多问题提供新的求解思路。
中文关键词: 特征表达;属性特征;深度学习;特征提取;深度架构
英文摘要: Deep attribute feature learning (DAFL) is a new research direction, and its main objective is to study a unified deep architecture for learning features covered from low-level to high-level. We will propose some innovative deep learning methods, models and the corresponding solutions in this project. And we will apply proposed methods to solve the problem of art retrieval by using abstract semantic attributes. The detailed content includes the analysis and estimation of single-modal deep learning method, the optimization of single-modal deep learning method and the study of learning higher semantic features. From the viewpoint of multi-modal feature fusion, we will study more valid deep learning architectures and optimization algorithms. By decomposing the attributes into a hierarchical structure, we will construct a map from multi-modal feature to attribute feature. We will test our proposed methods on several attribute data sets and give experimental results. Our work on deep attribute feature learning will not only promote the research on machine learning, but also provide new inspiration for many problems, such as large-scale feature learning, semantic annotation and content understanding, cross-media semantic retrieval, and so on.
英文关键词: feature representation;attribute featuer;deep learning;feature extraction;deep architecture