Empowering Semantic Indexing with Focus of Attention
Empowering Semantic Indexing with Focus of Attention
复制标题
DOI：
--
复制
发表时间：
2015
期刊：
影响因子：
--
通讯作者：
Kimiaki Shirahama;Tadashi Matsumura;M. Grzegorzek;K. Uehara
中科院分区：
文献类型：
--
作者：
Kimiaki Shirahama;Tadashi Matsumura;M. Grzegorzek;K. Uehara
本文讨论了语义索引（SIN）来检测视频镜头中的人和车等概念。一个主要的障碍是在同时显示多个概念的镜头中包含的丰富信息。换句话说，目标概念的检测受到在同一镜头中附带示出的其他概念的不利影响。我们假设用户可以识别目标概念时，它出现在一个显著的区域，吸引他/她的注意力。在此基础上，我们介绍了一种SIN方法，利用注意力的焦点（FoA）提取一个镜头中的显着区域，并构造一个功能强调该显着区域。此外，我们还开发了一种弱监督学习（WSL）方法来有效地为FoA创建训练镜头，以及一种镜头过滤方法来检查显着区域的有用性。实验结果表明，我们的SIN方法使用FoA的有效性。关键词语义索引;注意力焦点;弱监督学习;显著区域的重复性。
This paper addresses Semantic INdexing (SIN) to detect concepts like Person and Car in video shots. One main obstacle is the abundant information contained in a shot where multiple concepts are displayed at the same time. In other words, the detection of a target concept is adversely affected by other concepts which are incidentally shown in the same shot. We assume that a user can recognise the target concept when it appears in a salient region which attracts his/her attention. Based on this, we introduce a SIN method which utilises Focus of Attention (FoA) to extract a salient region in a shot, and constructs a feature emphasising that salient region. In addition, we develop a Weakly Supervised Learning (WSL) method to efficiently create training shots for FoA, and a shot filtering method to examine the usefulness of salient regions. Experimental results show the effectiveness of our SIN method using FoA. Keywords–Semantic indexing; Focus of attention; Weakly supervised learning; Usefulness of salient regions.