一种基于时空分布特征的新闻字幕检测新算法

新闻字幕中包含有丰富的语义信息，是实现自动化视频检索、分析和理解的重要信息源之一。本文通过研究新闻字幕的时空分布特征，提出了一个新闻字幕检测新算法，该算法首先用有师学习的方法获取字幕大小、方向、形状等信息；然后根据视频字幕区域的灰度空间差分分布特性，利用字幕的尺寸约束检测字幕，根据字幕时间持续特点，去除时间冗余；最后对字幕区域进行二值化。实验表明该算法对非滚动标注字幕检测效率高，实时性好，且对不同新闻视频具有较强的自适应性。

A New Method of News Local-caption Extraction Based on Spatio-temporal Distribution Feature

News local-caption is one of the important sources for the realization of automatic retrieval, analysis and comprehension of videos since it implies lots of semantics inherently. A new method is presented to detect and locate news local-captions using spatio-temporal distribution feature of local-caption. Firstly, the size, direction and shape of local-caption are extracted using supervised learning method; then the local-caption district is detected and located by size restrict based on space distributing feature of gray difference, and its time-redundancy is wiped off accordingly by time-continuance model; finally the local-caption is binarized. The method is experimentally satisfying in both processing efficient and real-time for non-moving editing local-caption and is adaptive to various news videos.