摘要
新闻视频中存在的大量闪光灯镜头造成镜头检测正确率的下降。通过分析闪光发生前后图像内容的相关性,提出一种基于梯度距离的镜头检测算法。首先针对序列直方图差的计算,提出基于局部最小距离的计算方法,消除由于拍摄光线微弱变化对原本静止的图像造成的直方图变化;然后对序列中可能存在的镜头帧,根据其出现的时间间隔和直方图距离比例,使用两步自适应阈值法过滤掉可能存在的渐变或特技镜头;最后对可能误检的闪光镜头,提出一种基于图像梯度向量的过滤方法,将候选镜头帧与前后连续图像的梯度向量进行对比,过滤掉差值小于阈值的图像。实验结果显示,该算法对新闻视频中的闪光镜头有较好的过滤效果,对连续闪光也有一定的鲁棒性。
A large number of flash shots in news video cause the decrease in shot detection accuracy. Through analysing the correlation of images before and after the flashing, we propose a gradient distance-based shot detection algorithm. First, aiming at the sequence histogram difference calculation, we propose a minimum local distance-based calculation method which eliminates the histogram variation caused by faint changes of shooting light on the original still image; then we use 2-step adaptive thresholding method to filter the gradual shot transitions and effect shots may existed on the shot frames possibly existed in sequence according to the time interval of their occurrence and the histogram distance proportion ; Finally for the flash shots possibly detected in error, we present a image gradient vector-based filtering method, and com- pare the candidate shot frames with the gradient vectors of their consecutive frames ahead and after, filter those images whose differences are smaller than the threshold. Experimental results show that the algorithm has preferable filtering effect on flash shots in news video, and has certain robustness on continuous flash.
出处
《计算机应用与软件》
CSCD
北大核心
2013年第8期130-133,147,共5页
Computer Applications and Software
基金
人工智能四川省重点实验室开放基金项目(2011RYY01)
