项目名称: 基于多义性码书学习和主题建模的图像语义分类技术研究
项目编号: No.61273257
项目类型: 面上项目
立项/批准年度: 2013
项目学科: 自动化技术、计算机技术
项目作者: 杨育彬
作者单位: 南京大学
项目金额: 82万元
中文摘要: 本项目围绕基于多义性码书学习和主题建模的图像语义分类技术及其应用开展研究。在基于"词袋"的主题模型基本研究框架的基础上,借助于面向多义性对象的机器学习、概率图模型和概率主题模型等领域的最新研究成果,有效结合图像的各类上下文信息和图像多义性的表示和描述,提出并实现一种能有效结合图像的各类上下文信息和图像多义性的新型主题模型框架;结合多义性学习理论和自调式学习机制,提出并实现一种面向图像多义性的有监督码书学习算法;基于概率图模型对码字上下文关系进行建模,提出并实现一种融合码字上下文关系的码书编码方法;提出并实现一种能有效融合图像空间和特征空间中上下文关系的汇合方法;提出一种基于上下文和多义性的图像概率主题分类模型,最终实现一种有效的图像语义分类方法;实现一个基于Internet 的分布式图像语义分类原型系统,并结合互联网图像检索等实际问题对本项目研究成果进行实验验证和应用。
中文关键词: 多义性学习;主题模型;码书学习;图像分类;上下文关系
英文摘要: This project will study image classification technologies and applications integrating codebook learning and multiple-semantics topic modeling. Based on the topic modeling framework built on "Bag-of-Word" representation, this research will propose a novel topic modeling framework integrating versatile contextual information and multiple semantics in images by using the state-of-the-art methods in multiple-semantics machine learning, probabilistic graph models and probabilistic topic models. Afterwards, this research will propose a supervised codebook learning algorithm aiming to address multiple semantics in images, and further model the contextual relationships by using probabilistic graph models to implement a novel coding method integrating the contextual relationships. Furthermore, this research will propose a feature pooling algorithm combining the contextual relationships both in image space and feature space. Finally, after designing a novel probabilistic topic model accommodating contextual relationships and multiple semantics, this research will implement an effective image classification method. In order to validate the above research results, this project will also implement an Internet-based image classification prototype system with the concern of real applications such as Internet-based image retri
英文关键词: Multiple-Semantics Learning;Topic Model;Codebook Learning;Image Classification;Contextual Relationship