基于镜头的视频场景构造方法研究

由于内容颗粒度地小,镜头层次的检索不能满足视频内容使用的需要。场景比镜头高一个层次的视频内容结构单,能在一定程度上缓解镜头颗粒度过小的问题。“场景”是一组镜头的集合,在内容上包含相似的对象或包含类似的背景。本文提出了一种基于镜头构造频场景的思路,包括三个环节：镜头边界探测,镜头特征提取和镜头聚类。

CONSTRUCTION OF VIDEO SCENES BASED ON SHOTS

In this paper, we argue that shots is incapable of rea video content consumption because of its too smalt content granularity and we put forward the concept of "scene" to partially solve this problem. Scene is a higher video content structure unit than shot. A scene is a cluster of similar shots that focus on same object(s) or display same background. We propose a complete 3-step solution to construct video scenes based on shots, including shot boundary detection, shot feature extraction, and shots clustering. Specially chosen or designed algorithms for these three steps tuning towards the task of scene construction are described, They are twin histogram comparison for shot detection, HSV color histogram calculation for feature extraction, and self-adjusting iterative shot clustering. And an experimental result is appended and discussed.