级联手工特征与深度特征的视频关键帧检测方法

关键帧检测是有效的视频内容分析的关键环节。常用的基于手工特征的方法运行效率高但很难有效表征关键帧特征,因而性能不好。基于深度特征的方法因为网络结构复杂,导致效率不高。在体育比赛类视频中,关键帧常为比赛转播中镜头变化的最后一帧。但广播视频中除了包含比赛视频还包括很多其他类型的镜头如中场休息、渐变镜头等。因此检测最后一帧包含很多比赛无关内容。针对这一问题,本文提出了一种手工特征与深度特征相结合的视频关键帧检测方法。首先基于颜色直方图特征进行镜头边界检测获取最后一帧。进一步基于直方图相似性提出一种类似聚类的方法得到候选关键帧。最后,基于深度神经网络对候选关键帧进行分类,得到真正的关键帧。在冰壶比赛视频和篮球比赛视频上的对比实验结果表明,相对于传统的背景差分法、光流法等,本文提出方法能够快速、可靠地提取关键帧。 

Video Key Frame Detection Method by Cascaded Manual Feature and Depth Feature

Key frame detection is the key link of effective video content analysis. The commonly used methods based on manual features are efficient but difficult to represent key frame features effectively, so the performance is not good. Because of the complexity of network structure, the method based on depth feature is inefficient. In sports games video, the key frame is often the last frame of shot change in the game broadcast. However, in addition to the game video, there are many other types of shots in the broadcast video, such as halftime, gradient shot and so on. So the last frame contains a lot of irrelevant content. In order to solve this problem, this paper proposes a video key frame detection method which combines manual feature and depth feature. Firstly, the last frame is obtained by shot boundary detection based on color histogram feature. Furthermore, based on histogram similarity, a similar clustering method is proposed to get candidate keyframes. Finally, the candidate keyframes are classified based on the depth neural network to get the real keyframes. The experimental results on curling match video and basketball match video show that compared with the traditional background difference method, optical flow method, etc, This method can extract key frames quickly and reliably.