视频质量评价研究综述  

A Survey on Recent Advances in Video Quality Assessment

作　　者：鄢杰斌 方玉明[1] 刘学林 姚怡茹 眭相杰 YAN Jie-Bin;FANG Yu-Ming;LIU Xue-Lin;YAO Yi-Ru;SUI Xiang-Jie(School of Information Technology,Jiangxi University of Finance and Economics,Nanchang 330013)

机构地区：[1]江西财经大学信息管理学院,南昌330013

出　　处：《计算机学报》2023年第10期2196-2224,共29页Chinese Journal of Computers

基　　金：国家自然科学基金(No.62132006);中国博士后科学基金面上项目(No.2022M721417);江西省自然科学基金青年项目(No.20224BAB212012);江西省教育厅科技项目一般项目(GJJ2200522)资助.

摘　　要：移动互联网时代每天都产生海量的质量参差不齐的视频数据,根据视频质量高效地过滤低质量视频对缓解设备存储压力起着至关重要的作用.此外,在视频的生成、处理、传输等过程中都不可避免地引入信号噪声,如何准确地预测视频质量,从而指导与监督视频处理与传输系统的优化具有重要的研究意义和实际价值.因此,视频质量评价受到越来越多的关注.视频质量评价旨在定量描述视频的视觉质量,包括主观质量评价和客观质量评价.主观质量评价通过开展视觉感知主观实验,研究各项因素对视觉质量的影响,并收集主观质量分数用于构建基准数据集;客观质量评价通过设计客观算法,自动预测视频的质量.本文首先介绍视频质量评价的基础知识,阐述视频质量评价的相关应用和问题;其次,重点介绍视频质量评价近二十年的发展现状,对比不同主观数据集的特点;然后,深入解析客观模型的建模思想,分层次对比不同的模型,详细分析各模型的优缺点;最后,指出未来发展方向并总结全文.In the current era of mobile internet,massive videos with varied quality have been produced daily.Therefore,it is of great importance to screen out low-quality videos quickly according to the predicted video quality to effectively relieve the storage pressure.In addition,distortion may be introduced inevitably in the procedure of video production,processing,transmission,display,and etc.Thus,estimating video quality accurately can be used for system optimization and algorithm optimization.Due to the above applications,video quality assessment(VQA)has gained more and more attention from both academia and industry.VQA aims to describe the quality of videos quantitatively,and it includes subjective quality assessment and objective quality assessment.The former means conducting psychophysical experiments,by which we can deeply explore the influence of different variables on video quality and collect subjective ratings for building benchmarking datasets,and the qualitative results of psychophysical experiments are often regarded as the guidance of designing objective VQA models.There are many mature and commonly used standards regarding collecting subjective ratings,such as single stimulus continuous quality evaluation and pair comparison,followed by outlier removal and final quality score collection.The latter means designing objective VQA models for automatically and accurately predicting the quality of videos.According to the accessibility of reference information,VQA models could be classified into three categories,including full-reference(FR),reducedreference(RR),and no-reference(NR).FR-and RR-VQA models need complete and partial reference information respectively when being deployed,and they often obey the‘similarity’measurement paradigm,i.e.,the video which is more similar to the associated reference video is regarded to be of better visual quality.By contrast,NR-VQA models can predict the quality of videos without access to reference information,and they usually follow the‘feature extraction and quality r

关 键 词：视频质量评价 视觉感知 特征工程 机器学习 深度学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]