基于3D DenseNet的视频镜头边界检测方法  

Vedio shot boundary detection method based on 3D denseNet

作　　者：张翔[1] 赵晓丽[1] 张嘉祺 陈正 张佳颖 王国中 ZHANG Xiang;ZHAO Xiao-li;ZHANG Jia-qi;CHEN Zheng;ZHANG Jia-ying;WXNG Guo-zhong(Shanghai Univerity of Engineering Science,Shanhai 201600,China)

机构地区：[1]上海工程技术大学

出　　处：《光电子．激光》2019年第10期1103-1109,共7页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金项目(61772328)

摘　　要：视频镜头边界检测(shot boundary detection,SBD)是视频检索中的关键预处理步骤,视频的每个帧段都全被归类为渐交、切变或不变。针对渐变镜头检测难度高和计算量过大的问题,本文提出了一种基于3D卷积和DenseNet相结合的深度镜头边界检测(DSBD)算法。该算法首先将视频分帧段随机分配标签,输入到3D DeseNet网络中,将具有相间标签的输出帧段合并,然后使用颜色直方图法测量帧段之间的巴氏距离来进行二次分类,最后能输出正确的帧段。通过在目前最常用的数据集UCF101_SBD和TRECVID以及最大的镜头检测数据集ClipShots上实验表明,该方法具有良好的检测效果,且计算时间较短,优于之前的算法。Video shot boundary detection(SBD) is a crucial pre-processing step in video retrieval,each segment of frames is classified as either sharp,gradual or no transition.The difficulty of gradual shot detection and huge amount of calculation has always been difficult problems to solve.Aiming at the problems of highdifficulty of gradual shot detection and huge amount of calculation,we present a doop shot boundary detection on(DSBD) algorithm based on 3 D convolution and DenseNet.Firstly,we divide videos into the segments of frames and give random labels.Then we input the segments into 3 D DenseNet and merge the output segments with the same label.We use color histogram to measure the Bhattacharyya distance between these segments.Finally,we output the correct segments.By doing experiments on the most frequently used datasets UCF101_SBD,TRECVID and largest datasets ClipShots currently,they show that this method has a good detection effect and shorter calculation time,which is better than the previous algorithms.

关 键 词：镜头边界检测 3D卷积 DenseNet 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]