项目名称: 交互式图像搜索中的小样本学习问题研究
项目编号: No.61301185
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 无线电电子学、电信技术
项目作者: 邬俊
作者单位: 北京交通大学
项目金额: 30万元
中文摘要: 小样本学习问题是交互式图像搜索技术面临的主要难题之一。现有应对方案主要致力于单种学习算法的设计与分析,却忽略了多种学习算法之间的交互与协同,其泛化性能尚存在较大改进空间。鉴于此,本项目主张以"完形认知机理"引导"机器学习"算法设计的学术思路,拟探索半监督学习、集成学习、主动学习,以及基于日志的长期学习等多种算法之间的协作机制,旨在构建基于多策略协同学习框架的交互式图像搜索新范式,以有效应对小样本学习困难;重点研究基于个性化用户行为模式挖掘的语义相似性推理方法、基于半监督集成理论的视觉相似性学习方法、融合语义与视觉信息的综合相关性计算方法、兼顾图像样本信息量与视觉显著性的混合型主动学习方法,以期最大化学习系统的泛化性能、最优化目标图像的相关性排序、最小化用户的标注负担。该项目的实施将进一步丰富和完善图像搜索研究领域的新内容,其科学、社会和经济意义重大,有望取得一些创新性成果。
中文关键词: 图像检索;机器学习;小样本;;
英文摘要: Interactive image search is often challenged by the small sample problem. Existing schemes mainly focus on designing and analyzing a single learning mechanism and ignore the cooperation among multiple learning strategies, which leads to very limited learning effectiveness. This project employs Gestalt laws to guide machine learning algorithms design, and proposes a novel interactive image search paradigm based on the multi-strategies cooperative learning framework to address the small sample learning problem, by developing the co-working mechanism of multiple learning approaches including semi-supervised learning, ensemble learning, active learning and log based long-term learning, etc. Our research mainly concentrates on semantic similarity inference by mining individualized behavior patterns of users, visual similarity learning using semi-ensemble theory, comprehensive relevance computation via combining image visual content with semantic hidden in query log, and hybrid active learning by querying both informativeness and visual saliency of image samples, in order to maximize the generalization performance of the learning system, optimize the relevance ranking of database images and minimize the user labeling burden. The implementation of this project will further enrich and advance the research of image retri
英文关键词: Image Retrieval;Machine Learning;Small Sample;;