项目名称: 基于相似性的图像特征逆向学习算法与应用
项目编号: No.61300072
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 自动化技术、计算机技术
项目作者: 刘一
作者单位: 北京交通大学
项目金额: 26万元
中文摘要: 设计图像特征是计算机视觉领域内的最基本研究内容之一,优秀的图像特征能够有效地提高图像分类、识别与检索等相关算法的性能。常见的图像特征分为全局和局部特征两大类,通常,这些特征都是利用图像中像素点的信息计算得到。然而,人为设计的图像特征之间的差异往往不能足够理想地反映图像块的相似性,因此这些正向设计的特征往往在某个图像处理任务上有较好的性能,但不能保证较好的通用性。本研究的核心是从图像中的图像块间相似性度量出发,逆向生成图像块的特征。这种特征构造方法是传统上由特征计算相似性的逆过程,其优点是图像块之间的特征差异能够理想地反映图像块之间的相似性并易于分类器设计。本项目主要研究内容包括:图像块之间的相似性度量;根据该相似性度量设计自动的特征提取算法以及图像间相似性度量的构造。其最核心的问题是如何逆向构造图像特征,我们将设计基于矩阵不完整Cholesky分解的算法完成这个关键的步骤,并证明其有效性。
中文关键词: 图像特征;相似性;高效核描述子;Cholesky分解;监督学习
英文摘要: Designing good image features in general or for a specific task is a longstanding research problem and a grand challenge for the computer vision community. And it is known that excellent image features are esential for boosting the performance of scene categorization, object recognition and image retrieval systems. Up to now, most common image features are either global or local ones that are computed in a pre-defined manner based on the information of image pixels. However, similarities between these manually curated image features usually do not represent the true affinities between these image patches. As a result, those "forward designed" image features may achieve good performance on some specialized image processing tasks, but they are not ideal candidates for more general computer vision problems such as natural scene categorization and retrieval. To resolve this problem, in this proposal, we investigate a novel approach to reverse engineering features from similarities between image patches. Different from traditional "forward" feature design approaches, the proposed approach has the unique advantage that the similarities between those reversed image features could represent the real affinities between image patches arbitrarily well, which also makes the classifier design much easier. In summary, the ke
英文关键词: Feature of an image;Similarity;Efficient Kernel Descriptor-EKD;Cholesky decomposition;Supervised learning