A new methodology to estimate the impact of H.264 artefacts on subjective video quality

A new methodology to estimate the impact of H.264 artefacts on

subjective video quality

Stéphane Péchard, Patrick Le Callet, Mathieu Carnec, Dominique Barba

Université de Nantes – IRCCyN laboratory – IVC teamPolytech’Nantes, rue Christian Pauc, 44306 Nantes, France

Third International Workshop on Video Processing and Quality Metrics for Consumer ElectronicsScottsdale, 2007-01-26

2

Introduction

• Codec => Coding artefacts

• Quality loss due to artefacts

=> Useful for quality metrics or better coding …

• Possible practical approach– Artefact classification– Annoyance or quality loss contribution per artefact type

3

Farias and al. methodology Farias VPQM 05

• Artefacts type set (blockiness, blur, ringing, …)• Generation of synthetic artefacts

– Strength parameter– Applied with the same strength on a whole part of the

sequence• Subjective assessment => Annoyance curve per artefact type regarding the

strength => Content dependency

Alternative approach : VPQM07⇒ H.264 coding, Subjective assessment : quality scale, blur

scale, blockiness scale …⇒ No direct control of artefacts strength

4

Proposed approach

• H.264 artefacts due to quantization/decision • effects are different regarding the local content

(edge, texture, …)• different perceived annoyance depending on the

local spatio-temporal activity of the content

• H264 distortions only in selected coherent spatio-temporal regions => define content categories

• Subjective quality assessment⇒ Quality loss curve per local content category

(e.g. effects of H264 on each category)

⇒ Strength ?

5

Outline

• Spatio temporal segmentation • distorted sequences generation• subjective quality assessment of sequences• Quality assessment : Combining categories• Towards quality loss function per content

category

6

The approach

temporalsegmentation classification

H.264 codingC-distortedsequencesgeneration

unlabeledholes filling

bordersprocessing

source

categories masks sequence

partly-distorted sequencesusable for subjective tests

Ci

……

7

Spatio temporal classification

2 steps

- temporal segmentation :reliability regarding the motion => temporal tubes

- tube classification :Regarding spatial content

8

Segmentation of sequences

temporalsegmentation classification

H.264 codingClass-distorted

sequencesgeneration

unlabeledholes filling

bordersprocessing

source

partly-distorted sequencesusable for subjective tests

Ci

……

9

Segmentation of sequences

• per group of five successive frame, the center frame is divided into blocks

• motion estimation of each block using the two previous frames and the two next frames

(motion estimation performed on a multi-resolution representation)

i+1 i+2i-1 ii-2

10

Segmentation of sequences

• temporal tracking of each block of frame i defines a spatio-temporal “tube” over the five frames

• a tube is oriented along the local motion

i+1 i+2i-1 ii-2

11

Classification

temporalsegmentation classification

H.264 codingC-distortedsequencesgeneration

unlabeledholes filling

bordersprocessing

source

partly-distorted sequencesusable for subjective tests

Ci

……

12

Definition of content categories

• low luminance smooth areas;• high luminance smooth areas;• fine textured areas;• edges;• strong textured areas

HVS has different perception of impairments depending on the local spatio-temporal content.

13

Classification

• 4 spatial gradients means per tube (directions : 0, 90, 45 and 135°)

• plot in spatial space P (0 and 90°) => C1, C2, C3 and C4

• 2nd step : space P’ (45 and 135°) used to discriminate C5 in P

• frontier determined to obtain relevant classification

14

Classification

• global tracking of moving objects over the whole sequence

• tubes are classified then merged by categories

smooth areas with low luminancesmooth areas with high luminancefine textured areasedgesstrong textured areas

15

Unlabeled holes filling and tube intersections

temporalsegmentation classification

H.264 codingC-distortedsequencesgeneration

unlabeledholes filling

bordersprocessing

source

partly-distorted sequencesusable for subjective tests

Ci

……

16

Unlabeled holes filling and tube intersection

• every pixel of the source has one and only one label• unlabeled holes :

– gradient value => class– closest tube

• Insection pixels : same

17

Borders processing

temporalsegmentation classification

H.264 codingC-distortedsequencesgeneration

unlabeledholes filling

bordersprocessing

source

partly-distorted sequencesusable for subjective tests

Ci

……

18

Borders processing

• borders between original and distorted large regions are treated so as to smooth the transitions

beforeborders

processing

afterbordersprocessing

19

H.264 coding and class-distorted sequences generation

temporalsegmentation classification

H.264 codingC-distortedsequencesgeneration

unlabeledholes filling

bordersprocessing

source

category masks sequence

partly-distorted sequencesusable for subjective tests

Ci

……

20

Partly-distorted sequences generation

H.264 sequences at different bitrates

categories sequence

original sequence C1

C2

C3

C4

C5

5 sequences per bitrate

21

Original sequence (first frame)

22

One caregory distorted sequence (first frame)

23

Subjective quality assessment

• SAMVIQ protocol with at least 15 validated observers and normalized conditions

• 1920x1080 HDTV Philips LCD display

• Doremi V1-UHD 1080i HDTV player

24

Subjective quality assessment

• 11 sequences in a SAMVIQ session:– 5 Ci-only distorted at a certain bitrate B

– entirely distorted sequence at B– entirely distorted sequence at low bitrate– entirely distorted sequence at intermediate

bitrate– explicit and hidden references

25

Sequences uncompressed HDTV sequences from SVT

Above marathon Captain Dance in the woods Duck fly

C5 50 % C2 78 % C3 54 % C5 60 %

Fountain man Group disorder Rendezvous Ulriksdals

C2 71 % C2+C3+C1 95 % C5 56 % C2+C3 80 %

26

example on sequence Ulriksdalscoded at 1 Mbps

0

10

20

30

40

50

60

70

80

90

1 2 3 4 5

Classes MOS(Sj,Bk) MOSref

27

• ∆MOS(Ci, Sj ,Bk) = MOSref - MOS(Ci, Sj ,Bk) is the quality loss induced by distortions in category Ci

∆MOS(C4)

∆MOS(C5)

∆MOS(C3)

∆MOS(C1)

∆MOS(C2)

MOSref

MOS(Sj,Bk)

DMOS(Sj,Bk)

MOS(C4)

MOS(C5)

MOS(C3)MOS(C1)

MOS(C2)

• MOS(Ci, Sj ,Bk) for each sequence Sj, each category Ci at each bitrate Bk

• DMOS(Sj ,Bk) = MOSref – MOS(Sj ,Bk) is the quality difference between the reference and the entirely distorted sequence

DMOS and ∆MOS

28

• relations use sums of ∆MOS

0.5472∆MOS(C4)

0.7664∆MOS(C2)

0.7094∆MOS(C3)

0.6400∆MOS(C5)

……

0.5349∆MOS(C1)

0.9058∆MOS(C1) + ∆MOS(C2) + ∆MOS(C3)

+ ∆MOS(C4) + ∆MOS(C5)

0.9094∆MOS(C2) + ∆MOS(C3) + ∆MOS(C4)0.9440∆MOS(C2) + ∆MOS(C5)0.9485∆MOS(C2)+ ∆MOS(C4) + ∆MOS(C5)

CCCombination

Possible relation between global DMOS and category ∆MOS?

29

Non linear functions

• DMOSp = maxi(∆MOSi)

– CC = 0.9467• DMOSp = maxi(∆MOSi) + maxj(∆MOSj) with j≠i

– CC = 0.9530

• Correlation exists between global DMOS and category ∆MOS=> DMOS could be predicted from quality per

category

30

Towards a quality loss model

• How to control the distortion level of a given class ?– Farias approach :strength of synthetic artefact

• Factors implied in the quality loss of category Ci:– distortions themselves– motion– proportion of the category– spatial localisation (not considered here)

31

Distortion strength for category C1

• distortion strength = f(M,P,E)With all along the sequence :– M the mean motion of the category;– P the mean proportion of the category;– E the MSE on the category;

• M decreases the distortion strength while P and E increase DSproposed model for f

DS = (1 — M/Mt)×P×E

32

Quality loss function for category C1

• Psychometic function as a prediction of ∆MOS1

φ(DS) = (a×DSb)/(c+DSb)

• correlation between φ(DS) and ∆MOS1 : 0.9514

• RMSE = 5.25• good predictor of the loss of quality induced by

category C1

33

Quality loss function for class C1

=> Possible prediction of ∆MOS1

34

Conclusion

• design of a new methodology to estimate the impact of H.264 artefacts on subjective video quality

• One distortion type but– Effect related to local content– possibility to relate the global loss to loss per

category– quality loss function for category C1

• Other categories and objective models

35

Results: segmentation statistics

60.7016.750.3650.06C5 (%)

10.703.021.430.94C4 (%)

19.5053.856.8127.79C3 (%)

8.9722.5778.2617.45C2 (%)

0.133.8013.143.75C1 (%)

Duck flyDance in the woodsCaptainAbove marathon

Séquence

36

Results: segmentation statistics

3.3056.924.543.93C5 (%)

1.362.051.791.45C4 (%)

40.4819.8729.8013.37C3 (%)

41.3112.3838.5870.71C2 (%)

13.548.7825.2810.52C1 (%)

UlriksdalsRendezvousGroup disorderFountain man

Séquence