Mining based on latent link analysis for large-scale multimedia archive

本文研究了大规模广播视频档案和网络图像视频等海量多媒体档案中可能对应的重复对象的局部图像检测与识别问题。我们将检测和识别的部分图像作为潜在链接，研究了从档案中获得的潜在链接组成的图结构的挖掘技术。该项目的成果实现了多媒体档案中出现的特定对象的枚举及其统计分析、基于元数据和注释共现分析的对象自动注释、自学对象识别系统等。部分图像检测和识别作为最重要的技术组成部分，将通过扩展我们之前的图像/视频复制检测和挖掘技术的研究成果来实现。

We address partial image detection and identification which may correspond to recurrent objects in huge amount of multimedia archives such as large-scale broadcast video archive and Web images and videos. We regard detected and identified partial images as latent link and study mining techniques for graph structure composed of latent links obtained from archives. The outcome of this project enables enumeration of specific objects appear in multimedia archive and their statistical analysis, object auto-annotation based on co-occurrence analysis with metadata and annotation, self-taught object recognition system, and so on. Partial image detection and identification, as the most important technical component, will be implemented by extending our former research outcome of image/video copy detection and mining techniques.