基于特征池与特征选择的低存储二值特征描述方法研究

项目名称： 基于特征池与特征选择的低存储二值特征描述方法研究

项目编号： No.61472119

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 刘红敏

作者单位： 河南理工大学

项目金额： 81万元

中文摘要： 局部图像特征描述子已经成功应用至许多计算机视觉系统，然而，现有性能较好的浮点型特征描述子计算复杂度高、存储空间要求大，不适合实时性要求高以及数据量大的应用。因此，为满足大数据与移动互联网时代背景下视觉应用对局部图像特征描述的需求，特征描述子不仅要具有很强的区分能力和鲁棒性，而且要存储开销小、计算和匹配的复杂度低。基于此，本项目研究快速、鲁棒、低存储的局部图像二值特征描述子，首先基于浮点型特征描述子的成功设计经验和二值特征描述子的特点，构建出一个区分能力强、计算复杂度低的由海量二值特征组成的特征池；然后基于机器学习中的特征选择技术从该特征池中选择合适的二值特征，研究数据驱动的基于特征选择技术的二值特征描述子构造方法；其次，研究基于Hash映射与稀疏表示的二值特征描述子的快速构造方法；最后，研究二值特征的浅层与深层神经网络融合方法，得到结构紧凑、区分力强的二值特征描述子.

中文关键词： 二值特征池；特征选择；特征描述；特征融合；特征学习

英文摘要： Local image descriptor has been successfully used in many applications. However, the currently top performance descriptors are floating-point descriptors, which require a large memory footprint and slow to compute and match. These drawbacks make them not applicable to real time and big data applications. To fulfill the requirement of vision applications in the age of big data and mobile internet, local image descriptor is expected to be not only with high discriminative ability and robustness, but also with low memory footprint and computational complexity. To address this problem, this proposal aims to develop local binary descriptors with high discriminative ability and robustness, low memory footprint and computational complexity. First, based on the experience of designing floating-point descriptors and characteristic of local binary descriptor, a binary feature pool is constructed by a large amount of binary features with high discriminative ability and computational efficiency; then based on the feature selection techniques, how to effectively construct data-driven local binary descriptors is studied in this project. Second, research on designing local binary descriptors by effective and efficient Hash mappings. Finally, research on how to effectively fuse different binary descriptors by shallow and deep network respectively, so as to obtain fused binary descriptors with high performance and very low dimensionality.

英文关键词： binary pattern pool;pattern selection;pattern description;pattern fusion;pattern learning