一种基于镜头切换和文本检测的广告检测算法

针对广告的两个基本特征： 镜头切换较频繁和经常突显商标信息, 提出一种基于镜头切换和文本检测的广告检测算法. 该算法先对各相邻帧进行直方图差分, 再对剪切、 消融、 淡入淡出和划变4种常见的镜头切换方式进行分析和识别, 实现了镜头切换检测. 对于镜头切换频率较低的广告, 通过基于最大梯度差分的文本检测方法检测广告商标信息, 实现了广告自动检测系统. 实验结果表明, 对于约15 min的视频序列, 该检测算法可获得83.32%的正确率, 可初步满足广告检测的要求.

TV Advertisement Retrieval Based on Shot Change and Text Detection

Based on the two basic features of advertisements: frequent shot changes and apparent brand information, the curve of histogram difference was obtained via histogram difference between two adjacent frames, and four common shot changes such as cut, dissolve, fade\|in/out and wipe were analyzed so as to recognize shot change detection. As for those advertisements with low shot change frequencies, the text detection method based on maximum gradient difference was employed to extract the brand information. According to the above ideals and algorithms, a video advertisement retrieval system was designed and realized. A general precision rate of 8332% in about 15 min of video frames indicates the system could preliminarily satisfy the performance requirement according to video consecutive sequence.