A Research on Extraction and Application of Video Keyword based-on Large-Scale Video Archive

提出了一种提取视频中的特征区域--“视频关键词”的方法，并对其应用进行了研究。视频关键词可以在视频中相互识别，从而实现基于特征对象、建筑物、人物、背景中的特定区域和图标区域的视频索引，这是传统的基于全局颜色分布索引的图像检索方法无法处理的。视频关键词提取一般需要图像分割来定位目标区域，但据说没有图像内容的语义信息就不可能实现完美的分割，这是不能假设的。因此，视频关键词提取可能是一项艰巨的任务。另一方面，基于每个关键词在TFIDF等大规模文本语料库中出现的统计信息，稳健地实现了用于文本分析的关键词提取，以选择重要的词，即关键词提取不依赖于语义信息。本研究的基本思想是将类似的技术应用到视频领域，即假设频繁使用的视频片段对应于重要的视觉信息，将其提取为视频关键词，显然需要人脸图像匹配才能实现从大规模视频档案中提取视频关键词。为此，我们提出了一种基于亮度分布直方图的优化特征表示方法--广义直方图，并对其有效性进行了验证。此外，还开发了使用相同视频片段检测来提取视频关键词的图像浏览器，以显示其对视频关键词的实际有效性。此外，为了进一步提高图像匹配的速度，提出了准高维聚类方法。

We proposed a technique to extract "video keywords" which are characteristic regions in videos, and we studied its applications. Video keywords can be identified each other in videos, and thus they enable video indexing based on characteristic objects, buildings, people, particular regions in background, and iconic regions, which ordinary approaches for typical image retrieval using indexing by global color distribution cannot handle.Video keyword extraction requires image segmentation to locate object region in general, however, it is said that the perfect segmentation is impossible without semantics information of the contents of the image, which cannot be assumed. Therefore video keyword extraction can be hard task. On the other hand, keyword extraction for text analysis is robustly realized based on statistical information of the occurrence of each keyword in large-scale text corpus, such as TFIDF, to select important words, i.e., keyword extraction does not rely on semantics information. Basic idea of the presented research is to apply similar technique to video domain, i.e., extract frequently used video segments as video keywords assuming that they correspond to important visual information.It obviously requires face image matching to realize video keyword extraction from large-scale video archives. So we proposed Generalized Histogram as an optimized feature representation based on luminosity distribution histogram, and its effectiveness was demonstrated. Moreover, the image browser using identical video segment detection to extract video keywords was developed to show its practical effectiveness of video keywords. In addition, the quasi high-dimensional clustering method was developed to further speed up the image matching.