基于深度学习的视频质量评价方法研究综述  

Literature Summary of Video Quality Assessment Methods Based on Deep Learning

作　　者：杨文兵 邱天 张志鹏 施博凯 张明威 YANG Wenbing;QIU Tian;ZHANG Zhipeng;SHI Bokai;ZHANG Mingwei(Joint Laboratory of Digital Optical Chip of Wuyi University and Institute of Semiconductor Research,Chinese Academy of Sciences,Jiangmen 529020,China)

机构地区：[1]五邑大学中国科学院半导体研究所数字光芯片联合实验室,广东江门529020

出　　处：《现代信息科技》2024年第7期73-80,85,共9页Modern Information Technology

基　　金：2021年江门市创新实践博士后课题研究资助项目(JMBSH2021B04);广东省重点领域研发计划(2020B0101030002)。

摘　　要：互联网时代充斥着海量的质量参差不齐的视频,低质量的视频极大地削弱人的视觉感官体验同时对储存设备造成极大压力,进行视频质量评价(VQA)势在必行。深度学习理论的发展为视频质量评价提供了新的思路,首先简单介绍视频质量评价理论知识和传统的评价方法,其次对基于深度学习的评价模型进行神经网络分类——2D-CNN和3D-CNN,并分析模型的优缺点,再次在公开数据集上分析经典模型的性能表现,最后对该领域存在的缺点和不足进行总结,并展望未来的发展趋势。研究表明:公开的数据集仍不充足;无参考的评价方法最具发展潜力,但其在公开数据集上的性能表现一般,仍有很大的提升空间。The Internet era is full of a large number of videos with uneven quality.Low quality videos greatly weaken people's visual and sensory experience and cause great pressure on storage equipment.Therefore,Video Quality Assessment(VQA)is imperative.The development of Deep Learning theory provides a new idea for video quality evaluation,which is of great significance to video quality evaluation.Firstly,the theoretical knowledge of video quality evaluation and traditional evaluation methods are briefly introduced,and then the evaluation models based on Deep Learning are classified by neural network(2D-CNN and 3D-CNN),and the advantages and disadvantages of the models are analyzed.Then the performance of the classical models is analyzed on the open data set.Finally,the defects and deficiencies in this field are summarized,and the future development trend is forecasted.The research shows that the open data set is still insufficient,and the evaluation method without reference has the most potential for development,but its performance on the open data set is average,and there is still a lot of room for improvement.

关 键 词：深度学习 视频质量评价 2D-CNN 3D-CNN 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]