摘要
在全媒体时代,媒体内容的表现形式逐渐丰富,开始成为影响信息传播的一个重要因素。内容质量评价仍停留在“流量思维”阶段,难以客观评价内容质量,亟需发展以用户为中心的全媒体内容质量评价方法。本文主要概述近十年来国内外公开发表的不同媒介的评价模型,回顾了图像、视频、音频、文本四类的客观质量评价在全媒体数据中的研究工作及相应的应用,主要介绍基于传统方法和基于深度学习方法两大方向中一些影响力较大的方法,每类方法有分成有参考和无参考的方法,对此总结了各方法特点,对一些具有代表性的方法进行了实验对比分析。最后对四种媒介内容质量评价领域仍面临的问题进行了总结并展望未来可能的发展方向。
In the era of all media,the forms of media content are gradually enriched and become an important factor affecting the dissemination of information. Building models to assess the quality of omnimedia content has attracted increasing attention. Quality assessment methods are mainly divided into subjective models and objective models.Subjective methods aim to assess quality through human eyes and human senses. It requires a lot of manpower and material resources,and the evaluation process also takes a lot of time;therefore it is difficult to apply in practical application.Objective quality assessment simulates the human observation process,which can automatically predict quality input.This review mainly summarizes the evaluation models of different media published at home and abroad in the past ten years. Research work and corresponding applications in omnimedia data. We mainly list some influential methods in the two major directions based on traditional methods and methods based on deep learning. The quality assessment models of the video and audio parts are divided into traditional methods and deep learning-based models. Each type of model is divided into reference models and non-reference models. Compared with the methods with reference data,the performance of the non-reference method has some differences. However,the no-reference quality evaluation model has strong applicability because it does not need to rely on reference information,and has always been a research hotspot in the field of image quality evaluation. The image part is mainly developed by dividing the unreferenced quality assessment model into supervised learning and unsupervised learning. The unsupervised method does not require the support of manual scoring data,saves labor,and has good development prospects. The text quality assessment model is introduced from the two directions of automatic scoring system and text generation quality assessment. Finally,it is concluded that traditional or applied deep learning methods have their own characteristics. These methods are independent of each other and form their own systems. It also looks forward to the possible development direction of all-media content quality assessment in the future.
作者
颜成钢
孙垚棋
钟昊
朱晨薇
朱尊杰
郑博仑
周晓飞
YAN Chenggang;SUN Yaoqi;ZHONG Hao;ZHU Chenwei;ZHU Zunjie;ZHENG Bolun;ZHOU Xiaofei(School of Automation,Hangzhou Dianzi University,Hangzhou,Zhejiang 310018,China)
出处
《信号处理》
CSCD
北大核心
2022年第6期1111-1143,共33页
Journal of Signal Processing
基金
全媒体信息传播理论与基础服务技术研究,重点研发计划(SY2020YFB1406600)
国家自然科学基金(61931008,61671196,62071415,62001146,61701149,61801157,61971268,61901145,61901150,61972123)
浙江省自然科学基金(LR17F030006,Q 19F010030)。
关键词
全媒体
图像质量评价
视频质量评价
音频质量评价
文本质量评价
omnimedia
image quality evaluation
video quality evaluation
audio quality evaluation
text quality evaluation
