基于亮度方差与深度学习的视频渐变镜头边界检测算法研究  

Video gradient shot boundary detection algorithm based on brightness variance and deep learning

作　　者：徐敏[1] 申奥 司小朦 Xu Min;Shen Ao;Si Xiaomeng(Shanghai Polytechnic University)

机构地区：[1]上海第二工业大学,上海201209

出　　处：《现代电影技术》2025年第1期45-51,共7页Advanced Motion Picture Technology

基　　金：上海第二工业大学校级青年项目“结合老电影损伤特点的旧影像色彩增强技术研究”(61402278)。

摘　　要：镜头是视频索引、搜索浏览和老旧影像快速修复的基本单元,因此开发高效镜头分割算法,使用户得以迅速定位切换点至关重要。本文通过详细分析镜头串接的特点,根据镜头发生切换时亮度方差均值直方图所表现出的特性差异找到切换点,从而达到分割镜头的目的。针对各种类型的渐变检测,首先,通过计算相邻图像间的光流矢量来校正图像偏移,并依据矢量量化结果和图像熵的评估确定所有潜在的镜头边界;之后,跟踪超像素以提取图像的显著区域,同时结合预训练的AlexNet模型学习视频帧的特征,并训练得到渐变镜头边界检测模型;最后,利用直方图帧差法和光流法排除由闪光和高度运动引起的误检,完成渐变镜头的边界检测。实验结果表明,本文方法在检测渐变镜头方面表现出色,同时对运动场景及闪光灯造成的干扰具有较强的稳健性。Shot is the basic unit of video indexing,browsing,and prompt restoration of old images.Hence,it is very impor-tant to develop an efficient shots segmentation algorithm and quickly locate the switching point.Through careful analysis of the characteristics of shots'connection,the switching point is found according to the characteristics of the histogram of brightness variance mean,so as to achieve the purpose of shots segmentation.For various types of gradient detection,firstly,the image offset is corrected by calculating the optical flow vector between adjacent images,and all potential shots boundaries are determined according to the vector quantization results and image entropy evaluation.Then,the superpix-els are tracked to extract the significant areas of the image,and the features of the video frames are learned by combining the pre-trained AlexNet model,and the gradient shot boundary detection model is trained.Finally,the histogram frame dif-ference method and optical flow method are used to eliminate the false detection caused by flash and height motion,and the boundary detection of the gradient shots is completed.The experimental results show that the proposed method per-forms well in the detection of gradient shots,and has strong robustness to the interference caused by moving scenes and flash lights.

关 键 词：镜头分割 渐变检测 亮度方差 分块直方图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]