基于手工特征的视频哈希研究综述

Survey of Video Hash Research Based on Hand-craft Features

视频哈希是从视频中提取到的基于视觉内容的短小数字序列,在实际应用中,用视频哈希来表示视频,能降低视频的存储代价和视频相似计算的复杂度。目前,视频哈希已被广泛应用于拷贝检测、篡改取证、视频索引、视频检索等方面。近年,视频哈希研究取得许多重要进展,研究人员设计和开发出多种手工特征提取技术,并建立一系列视频哈希算法。本文将基于手工特征的视频哈希算法分为空域计算和时空域计算2个大类,其中基于空域计算的哈希算法又分为逐帧计算和关键帧计算2类,而基于时空域计算的哈希算法则分为正交变换、统计特征、视觉特征点、数据降维和其他技术5类。根据这些分类,本文先分析每类算法的代表性研究成果并总结其性能;然后介绍常用的哈希度量方法、性能评价指标和视频数据集;最后列出未来研究工作可重点关注的内容,包括面向篡改取证的视频哈希、基于深度学习的高效视频哈希和面向移动应用的轻量级视频哈希等。

Video hash is a short digital sequence based on visual content extracted from video. In practical applications, the strategy of representing a video by its video hash can reduce storage cost of video and complexity of calculating video similarity. Video hash is widely used in many applications, such as copy detection, tampering forensics, video indexing, and video retrieval, etc. In recent years, there are many progresses on video hashing research. Researchers have designed and developed some techniques of hand-craft feature extraction and presented many video hashing algorithms. This paper classifies the hand-craft features based video hashing algorithms into two categories: spatial computation, and spatial-temporal computation. Moreover, the video hashing algorithms based on spatial computation are further divided into two sub-categories: frame-by-frame calculation and key frame calculation. The video hashing algorithms based on spatial-temporal computation are further divided into five sub-categories: orthogonal transform, statistical feature, visual feature point, data dimensionality reduction and other technique. In this paper, some typical algorithms of each category are first described, and their performances are then summarized. Next, the common-used metrics of hash similarity, performance evaluation indices and video datasets are introduced. Finally, the future trend of research is presented, including video hashing for tampering forensics, efficient video hashing based on deep learning, lightweight video hashing for mobile applications, and so on.