多方式数据分析驱动的视频质量评估：实用指南

需积分: 10 196 浏览量更新于2024-07-17 收藏 5.7MB PDF 举报

《设计视频质量度量：多方式数据分析的方法》是一本深入探讨在数字视频领域广泛应用的客观质量评估标准的专著。作者Christian Keimel在书中提出了一个数据驱动的方法，针对视频编码质量的评估，强调了多方式数据分析在这一过程中的关键作用。该书是T-Labs系列通信服务的一部分，由Sebastian Möller、Axel Küpper和Alexander Raake共同编辑，该系列旨在为电信服务领域的专业人士提供最新研究成果。书中详细介绍了当前视频质量评估标准的现状，这些标准包括但不限于PSNR（峰值信噪比），SSIM（结构相似性指数）等常用指标，以及更为复杂的多维分析技术，如深度学习在视频质量预测中的应用。作者通过案例研究和实证分析，展示了如何利用多元数据源，如视频帧间差分、纹理特征、运动矢量分析等，来综合评估视频的主观体验，提升客观评价的准确性和一致性。此外，书中还涵盖了数据采集、预处理、特征提取和模型构建的完整流程，以及如何通过统计学和机器学习技术进行实时监控和优化。作者强调了跨学科合作的重要性，包括计算机视觉、信号处理、通信工程和数据科学等领域，以确保视频质量评估的全面性和有效性。读者可以从中收获对视频编码技术的理解，如何运用现代数据分析工具解决实际问题，以及如何在不断发展的技术环境中，更新和改进视频质量评估标准。《设计视频质量度量：多方式数据分析的方法》不仅适合从事视频编码与质量评估的专业人员，也对研究人员和工程师在多媒体通信、网络传输和用户体验优化方面提供了宝贵的参考。该书的出版信息包括ISBN号、DOI、电子版信息和版权声明，表明了作者对知识产权的尊重。《设计视频质量度量：多方式数据分析的方法》于2016年由Springer Science+Business Media Singapore出版，具有国际认可的Library of Congress Control Number，体现了其学术价值和权威性。

2 1 Introduction

practice often leads to a blurring between these two different groups of visual quality

metrics in everyday use, but one should be aware of the fundamental difference of

still images and video.

1.1 Motivation

Visual quality metrics should fulﬁl three basic requirements: they should be able to

provide a visual quality estimation similar to human observers, be properly validated

and lastly should be applicable regardless of the availability of an undistorted ref-

erence in order to be usable in real-life applications. Video quality metrics should

additionally also take the temporal nature of video and the resulting variation of

visual quality over time properly into account and thus avoid any temporal pooling.

Considering the video quality metrics proposed so far in the state-of-the-art, however,

these four requirements are often only met partially:

Temporal nature of video Many video quality metrics account for the temporal

nature of video by using pooling approaches as the Minkowski summation [21, 57,

186, 349, 353–355] or percentile pooling [27, 119, 193], while others include frames

in a small interval around the current frame into the prediction [21, 186, 286] or

use the differences to the preceding frame in the prediction [78, 216, 229, 230, 233,

238, 281, 345, 374, 378]. Only a small subset of all metrics consider the temporal

dimension of video without any signiﬁcant temporal pooling or at least consider a

relatively large interval of a few seconds [11, 138, 175, 206, 224, 336].

Validation Validation of visual quality metrics is mostly limited to a single data set

and only a small number of quality metrics are validated with multiple datasets [222,

238, 283, 306, 341, 378], but none of the video quality metrics are.

Reference availability Only a limited number of video quality metrics do not require

the undistorted reference [23–25, 57, 58, 70, 138, 181, 233, 317, 318, 374]. Compared

to the overall number of technology-agnostic video quality metrics in the-state-of-

the-art, only a small subset fulﬁls this requirement [57, 58, 70, 138, 233, 374], in

contrast to coding technology-speciﬁc metrics that in the majority fulﬁl this require-

ment [23–25, 181, 317, 318].

Prediction performance Using a Pearson correlation coefﬁcient above 0.9 between

the visual quality perceived by human observers and the prediction provided by the

video quality metrics as a criterion for the metrics’ prediction performance, only

a small number of technology-agnostic metrics reach this goal [27, 70, 175, 193,

238, 281, 345], in contrast to technology-speciﬁc metrics that due to the focus on a

speciﬁc technology are generally able to achieve the required prediction performance

[23–25, 317, 318].

Although some video quality metrics in the state-of-the-art fulﬁl a subset of the

four requirements, no metric fulﬁls all requirements in particular with respect to the

proper consideration of the temporal nature of video.

1.2 Contribution of This Book 3

1.2 Contribution of This Book

In this book, I therefore propose a data driven design methodology using multi-way

data analysis methods for the design of video quality metrics in order to address

these shortcomings in the state-of-the-art, especially with respect to the inadequate

consideration of video’s temporal domain so far. Moreover, I demonstrate that this

approach in combination with simple features and proper validation allows us to

easily design metrics that are able to fulﬁl all four requirements outlined in the

previous section.

Data driven design methodology Unlike the concept behind most video quality

metrics in the state-of-the-art, the design methodology proposed in this book is

purely data driven. It neither assumes nor needs a model of the human visual systems

but utilises only objectively measurable properties of the video sequences, the so-

called features. These features are then used in combination with visual quality

ratings gained from human observers to built prediction models for the visual quality

in a training phase. Hence, no a-priori relationship between certain features and

visual quality is assumed, but the relationships between features and visual quality

emerge during the training. The resulting models are then validated with independent,

unknown sequences in order to assess the prediction accuracy. This data driven

approach is widely used in chemometrics, a research area that often faces similar

challenges to visual quality assessment in building prediction models for estimating

the response of human perception to olfactory or gustatory stimuli.

In this book, I provide an detailed overview how a data driven design paradigm can

be used in the design of video quality metrics. The major advantage of this approach

is the s eparation of the feature extraction and model building process, allowing for

the use of any features, using any model building method. Hence, even if only a

partial knowledge of the inﬂuence of the video sequences’ properties on the visual

quality as perceived by human observers is available or some inﬂuence is maybe

only suspected, the resulting features representing these properties can still be used

in building a model, as now the overall combination of these features in a prediction

model is independent from a a-priori model describing the exact relationship between

the features.

Data driven design with multi-way data analysis Prediction models in the data

driven design can be built with many different methods, but in this book I focus on the

data analysis approach as commonly used in chemometrics. Data analysis methods

have already previously been used successfully in the design of video quality metrics

[229, 230, 233], but the employed methods required a matrix or two-way array

presentation of the information gained from the video sequences and thus temporal

pooling was necessary.

In this book, I extend this approach by utilising multi-way data analysis methods

that are able to handle the three-way nature of video without preceding pooling and

thus consider the complete variation within a video sequence in the model building.

In particular, I focus on data analysis methods that do not only consider the variation

4 1 Introduction

within the features but also the variation within the visual quality during the model

building process by introducing the multi-way partial least squares regression into

the data driven design of video quality metrics. Furthermore, I also provide a com-

prehensive overview of two-way and multi-way data analysis methods, especially in

the context of the design of video quality metrics.

High performance example metrics In order to evaluate if the data driven design

with multi-way data analysis proposed in this book addresses the shortcomings of

existing video quality metrics, I used the proposed approach to design two differ-

ent example metrics, each representing one of the major categories encountered in

state-of-the-art metrics. The ﬁrst example metric represents universal or technol-

ogy independent metrics and uses features that are extracted from the pixels con-

tained within the video sequences. The second example metric represents technology

speciﬁc metrics and in is based on features extracted from the bitstream of video

sequences encoded according to the ubiquitous H.264/AVC standard. Both example

metrics are no-reference metrics that do not require the undistorted video sequence

for their quality prediction. As the focus in this book is on the data driven design

approach, the t wo example metrics are not optimised manually towards speciﬁc test

conditions, but are built using only multi-way data analysis without ﬁne tuning.

The results in this book show that the proposed approach with multi-way data

analysis outperforms the combination of two-way data analysis methods and tempo-

ral pooling. Moreover, even though purely data driven, the designed example metrics

also outperform existing visual quality metrics in the state-of-the-art that are often

highly optimised towards certain properties of the human visual system. In particu-

lar, the example metrics also fulﬁl all requirements discussed in the previous section

unlike many metrics in the state-of-the-art. These results are based on a comprehen-

sive validation using four different data sets spanning a large content, format and

visual quality range, suggesting the validity of the results and trends in this book

beyond the used video sequences.

1.3 Outline

This book consists of three distinct main parts as illustrated in Fig. 1.1: the ﬁrst part

provides an introduction into video quality assessement, the second part discusses

data analysis methods, and the third part design of video quality metrics demonstrates

how the data driven design approach with data analysis can be used in the design

of video quality metrics. Following the three main parts, the Appendix provides

additional details and results, complementing the three main parts.

Part I—Video quality assessment In Part I, I focus on the question of what quality in

general and video quality in particular is and how video quality can then be measured,

both subjectively and objectively.

Following a review of existing quality concepts in the context of multimedia

applications and services, I introduce in Chap. 2 the deﬁnition of video quality as

1.3 Outline 5

Part I

Video quality assessment

Part III

Metric design

Part II

Data analysis

Chapter 6

Multi-way

data analysis

Chapter 1 Introduction

Chapter 7

Model building

considerations

Chapter 5

Two-way

data analysis

Chapter 4

Data analysis

approach

Chapter 3

Video quality

metrics

Chapter 9

Performance comparison

Chapter 8

Designing

video quality

metrics

Chapter 10 Conclusion

Chapter 2

Video quality

Fig. 1.1 Structure of this book

used in this book that is based on the general concept of Quality of Experience

(QoE). In light of this deﬁnition, methodologies and environments for subjective

video quality assessment are discussed in detail that allow us to determine this video

quality using human observers.

Based on this understanding of video quality and its assessment using human

observers, I extend in Chap. 3 the s cope of video quality assessment to video quality

metrics, allowing us to evaluate video quality without human observers. Firstly, I

deﬁne and justify the requirements for video quality metrics, and introduce a taxon-

omy of video quality metrics that allows us to categorise the different approaches to

the design of video quality metrics. Using both requirements and the taxonomy, I then

evaluate the state-of-the-art in video quality metrics and highlight the shortcomings

of existing concepts in the design of video quality metrics.

Part II—Data analysis Part II is focused on the data analysis methods employed in

the proposed data driven design approach to video quality metrics.

In Chap. 4, I provide an introduction to the general concept of model building with

data analysis, followed by a discussion of the notation necessary to describe video

sequences in an adequate form for the application of the data analysis methods, in

particular with respect to the three-way nature of video. The chapter concludes with

the description of the preprocessing that should be applied before t he data analysis

is performed.

These preliminaries are followed by an in-depth discussion of traditional two-way

data analysis methods in Chap. 5. After brieﬂy reviewing temporal pooling methods

6 1 Introduction

that map three-way data to two-way data and that are necessary before two-way

data analysis methods can be applied, I start the discussion on data analysis methods

with multiple linear regression (MLR), before reviewing the concept of component

models and latent variables. I then discuss two-way component-based data analysis

methods that can be used to extract latent variables from the data, the well-known

principal component regression (PCR) and the lesser-known partial least squares

regression (PLSR).

Using two-way data analysis methods as a foundation, I discuss in Chap. 6 ﬁrst how

these two-way methods can be applied to three-way data without temporal pooling,

before introducing the concept of multi-way component models. Based on the well-

established concepts of the Tucker3 and PARAFAC decompositions for multi-way

component models, I provide a detailed discussion of the three-way extension of

the PLSR, the trilinear PLSR, that is used in this book for building video quality

prediction models. The discussion of the data analysis methods is concluded in

Chap. 7 by addressing issues with respect to validation, model selection, component

selection and feature selection that need to be taken into account in the data driven

design approach with data analysis.

In the context of this book, the focus is on prediction performance, not complexity,

and the algorithms for the different data analysis methods discussed in Part II may

not necessarily be the computational most efﬁcient ones. Similarly, the aim in this

book is prediction, not explanation, and thus although all data analysis methods per

deﬁnition allow for an interpretation of the features with respect to their inﬂuence on

the predicted video quality, this option is not pursuit further in this book. Lastly, even

tough the data driven approach can be implemented with many different methods, the

focus in this book is on analysis, not learning and therefore alternative concepts to the

data analysis approach e.g. machine learning or neural networks are not discussed.

Part III—Design of video quality metrics In Part III, I demonstrate how the data

driven design approach with multi-way data analysis discussed in Part II can be used

to design video quality metrics with a high prediction performance, providing results

similar to the results gained in subjective quality assessment as described in Part I.

Chapter 8 introduces the two example metrics that are used in this book to evaluate

if the proposed data driven design approach with multi-way data analysis is suitable

for the design of video quality metrics. Firstly, I describe the technology-speciﬁc

example metric based on features extracted from the bitstream of video sequences

encoded according to the H.264/AVC standard, before describing the second example

metric that represents universal or technology independent metrics, using features

that are extracted from the pixels contained within the video sequences. Both example

metrics are no-reference metrics that do not require the undistorted video sequence

for t heir quality prediction. As the focus of this book is on design, not development,

the two example metrics are not optimised manually towards speciﬁc test conditions,

but are built using only multi-way data analysis without ﬁne tuning. Also I only focus

on features, not vision, using only features extracted from the video sequences and do

not rely on modelling the human visual system, following Winkler’s remark in [355]

that the human visual system is extremely complex and that our current knowledge

剩余244页未读，继续阅读

peter15121

粉丝: 0
资源: 4

多方式数据分析驱动的视频质量评估：实用指南

大数据HDPAmbariambari-metrics-collector-2.7.7.0-0.x86-64.rpm

metrics-graphite-3.1.5-API文档-中英对照版.zip

metrics-librato-4.0.1.11.jar

metrics-server-0.3.7.tar.gz

metrics-librato-4.1.2.3.jar

metrics-librato-2.2.0.4.jar

metrics-librato-3.0.1.RC2.jar

metrics-librato-2.2.0.1.jar

metrics-librato-4.0.1.10.jar

metrics-librato-4.0.1.6.jar

最新资源