视频信号中实时字幕信息的提取方法

Real-time text information extraction from videos

为了在视频图像中进行字幕信息的实时提取 ,提出了一套简捷而有效的方法。利用视频图像中文本的频率特性与空间连续性 ,采用改进的投影阈值分割方法对视频中的文本进行实时分割。针对视频字幕在时间上的冗余特性 ,提出了一个基于有限状态机的动态缓冲的模型 ,在提高分割的正确率的同时减小了识别运算量。在识别部分 ,采用了一个 3层前向神经网络进行实时的识别。该算法已经成功地应用于卡拉 OK MTV歌词字幕信息同步提取系统中。

A simple and effective method is presented for real time text segmentation and recognition in videos. The frequeny and spatial characteristics of the text are analyzed by a fast segmentation algorithm developed from the conventional threshold method. A dynamic buffering algorithm based on the Finite State Machine is used to eliminate the text's temporal redundancy and at the same time to correct segmentation errors. The recognition algorithm employs a 3 layer BP NN for real time recognition. The algorithms have been successfully applied to a system which automatically extracts lyrics from MTV Karaoke videos.