symme y
S
S
A icle
P edic ing Pe cep ual Quali y in In e ne Tele ision
Based on Unsupe ised Lea ning
Ja osla F nda 1,* , Jan Nedoma 2, Radek Ma inek 3and Michael F id ich 2
1Depa men o Quan i a i e Me hods and Economic In o ma ics, Facul y o Ope a ion and Economics o
T anspo and Communica ions, Uni e si y o Zilina, 010 26 Zilina, Slo akia
2Depa men o Telecommunica ions, Facul y o Elec ical Enginee ing and Compu e Science,
VSB—Technical Uni e si y o Os a a, 17. Lis opadu 15, 708 33 Os a a-Po uba, Czech Republic;
[email p o ec ed] (J.N.); [email p o ec ed] (M.F.)
3Depa men o Cybe ne ics and Biomedical Enginee ing, VSB—Technical Uni e si y o Os a a,
708 00 Os a a-Po uba, Czech Republic; [email p o ec ed]
*Co espondence: ja osla [email p o ec ed]
Recei ed: 13 Augus 2020; Accep ed: 16 Sep embe 2020; Published: 17 Sep embe 2020


Abs ac :
Quali y o se ice (QoS) and quali y o expe ience (QoE) a e wo majo concep s o he
quali y e alua ion o ideo se ices. QoS analyzes he echnical pe o mance o a ne wo k ansmission
chain (e.g., u iliza ion o packe loss a e). On he o he hand, subjec i e e alua ion (QoE) elies on
he obse e ’s opinion, so i canno p o ide ou pu in a o m o sco e immedia ely (ex ensi e ime
equi emen s). Al hough se e al well-known me hods o objec i e e alua ion exis ( ying o adop
psychological p inciples o he human isual sys em ia ma hema ical models), each o hem has i s
own a ing scale wi hou an exis ing symme ic con e sion o a s anda dized subjec i e ou pu like
MOS (mean opinion sco e), ypically ep esen ed by a i e-poin a ing scale. This makes i di icul
o ne wo k ope a o s o ecognize when hey ha e o apply esou ce ese a ion con ol mechanisms.
Fo his eason, we p opose an applica ion (classi ie ) ha de i a es he subjec i e end-use quali y
pe cep ion based on a sco e o objec i e assessmen and selec ed pa ame e s o each ideo sequence.
Ou model in eg a es he unique bene i s o unsupe ised lea ning and clus e ing echniques such as
o e i ing a oidance o small da ase equi emen s. In ac , mos o he published pape s a e based
on eg ession models o supe ised clus e ing. In his a icle, we also in es iga e he possibili y o a
g aphical SOM (sel -o ganizing map) ep esen a ion called a U-ma ix as a ea u e selec ion me hod.
Keywo ds: mapping unc ion; QoE; QoS; sel -o ganizing map; ideo quali y es ima ion
1. In oduc ion and Mo i a ion
S eaming media has become a popula echnology in ecen yea s and ep esen s a majo amoun
o he da a deli e ed ia in e ne p o ocol (IP)-based ne wo ks. Adop ion o he nex -gene a ion
ne wo k (NGN) concep allows us o ansmi no only da a bu also oice and ideo in hei digi al
o m ia ansmission chains o iginally p ojec ed o da a ans e (such as e-mail communica ion,
web b owsing, e c.).
Since mul imedia se ices (especially ideo) a e gene ally asymme ic ( he con en is deli e ed in
one di ec ion) eal- ime c i ical se ices, ansmission con ol p o ocol (TCP) had o be eplaced by a
connec ionless use da ag am p o ocol (UDP) wi h no gua an ee o da a deli e y (da a e ansmission
causes addi ional o e head ime in he communica ion p ocess). The UDP a oids he o e head ime o
he TCP handshake p ocedu e. On he o he hand, los da a can c ea e isual impai men s du ing
ideo decoding and playback.
Well-known ideo on demand (VoD) se ices, e.g., HBO Go, Ne lix, and YouTube, s ill use a
eliable connec ion es ablished by TCP, bu a e in ac no li e s eam b oadcas ing. They use bu e s o
Symme y 2020,12, 1535; doi:10.3390/sym12091535 www.mdpi.com/jou nal/symme y
Symme y 2020,12, 1535 2 o 16
p eload da a in o ese ed a ea o memo y; o ins ance, YouTube s o es app oxima ely 60 s o he
encoded ideo be o e playback, and Ne lix has a 240-s playback bu e .
I p o ide s o in e ne p o ocol ele ision (IPTV) se ices wan o become success ul compe i o s
o adi ional e es ial ideo b oadcas ing se ice companies, hey mus con ol and analyze he
se ice quali y hey p omo e. Because subjec i e es s equi e he pa icipa ion o many obse e s,
he pic u e quali y is calcula ed by objec i e ideo quali y me hods. Se e al well-known me hods
ha e been de ined o his, such as peak signal o noise a io (PSNR), ideo quali y me ic (VQM),
and s uc u al simila i y index (SSIM). PSNR is he oldes me hod bu p o ides as calcula ion o
esul s; SSIM and VQM e lec be e on end-use subjec i e ideo quali y pe cep ion [
1
,
2
]. Gene ally,
subjec i e e alua ion is ep esen ed in he o m o i e-poin scale, s anda dized by he In e na ional
Telecommunica ion Union (ITU), called he mean opinion sco e, whe e i e poin s s ands o he
highes pe cei ed quali y. Each o hese objec i e me ics uses i s own scale. The ela ionship be ween
he subjec i e and objec i e esul s has no ye been de ined exac ly, so i is no known how o co ec ly
in e p e objec i e esul s on he subjec i e MOS scale.
The possibili y o using subjec i e MOS da a measu ed o di e en QoS scena ios has been
explo ed o come up wi h a model allowing he p edic ion o QoE based on QoS esul s, as well as he
de i a ion o QoS pa ame e s o a gi en QoE equi emen . In o de o achie e his, i is essen ial o
unde s and how measu able pa ame e s e lec he quali y o se ice. QoS and QoE a e impo an o
bo h cus ome s (indi iduals and businesses) and se ice p o ide s, and hei p o ision should no
only be moni o ed, bu also encou aged and en o ced when needed.
IPTV, as a pa o iple play (a ma ke ing e m o o e ing da a, oice, and ideo se ices om
one se ice p o ide ), is unde he egula ions in nume ous coun ies. Na ional egula o y au ho i ies
need o p epa e quali a i e c i e ia (a se o measu ed pa ame e s and hei limi s), as well as o ha e
applica ions able o measu e cus ome sa is ac ion le els. QoS egula ion should be a pa o cus ome
p o ec ion, aiming o p o ide an objec i e quali y o deli e ed se ice compa ison be ween p o ide s.
In addi ion, gi en he impo ance o be e pic u e quali y wi h IPTV, i is in he company’s bes in e es
o p o ide a high-quali y se ice; o he wise, he cus ome will go elsewhe e. Ou c ea ed model
akes in o accoun he concep s o bo h quali y e alua ion echniques, which esul s in a unc ional
moni o ing sys em.
Fo ha eason, he key objec i e and mo i a ion behind his wo k is o combine he esul s
p o ided by he objec i e and subjec i e me hods. The p oposed sys em will be in a o m o a
so-called passi e moni o ing (nonin usi e) ool ha does no a ec he inspec ed ansmission chain
(e.g., packe s manipula ion) and will be based on he machine lea ning algo i hm. Machine lea ning
o e s da a analysis ha au oma es p edic ion model building. A selec ed algo i hm can o ganize la ge,
complex da ase s, and he aining p ocedu e is e y as in compa ison o adi ional solu ions based
on back-p opaga ion neu al ne wo ks. We also ied o inspec he clus e ing abili y o algo i hm o
ea u e selec ion. A Kohonen map allows us o iden i y and selec impo an a iables in he ea u e
spaces, which is ano he bene i o he clus e ing echnique, bu only a ew s udies ha e con ibu ed o
his knowledge. Ou applica ion can be ope a ed as an end-use ideo quali y pe cep ion es ima o
and can help wi h da a low p io i iza ion se ings in con en deli e y ne wo ks.
2. Rela ed Wo ks
Se e al s udies p opose mapping unc ions o ansla e objec i e esul s in o a subjec i e poin o
iew. The i s endea o o apply machine lea ning o his esea ch issue was called pseudo-subjec i e
quali y assessmen (PSQA) [
3
]. Mohamed and Rubino wo ked wi h pa ame e s such as packe loss
o bi a e as an inpu o neu al ne wo k modelling, ollowed by he objec i e sco e, o compu e a
subjec i ely pe cei ed quali y. They used a e y old (nowadays) codec, MPEG 2, al oge he wi h low
esolu ion (352
×
288) and a small sample size da ase , bu hei a icle s ill se es as a basic e e ence
o his wo k.
Symme y 2020,12, 1535 3 o 16
Valde ama and G
ó
mez [
4
] chose a di e en se o inpu s, including di e en leng hs o he g oup
o pic u es (GOP), wo p io i iza ion echniques (Di Se o Bes E o ), and bandwid h bo lenecks in
he expe imen al ne wo k. They ob ained a Pea son co ela ion coe icien (PCC) sligh ly abo e 0.9,
bu only one esolu ion (740
×
480) and a high packe loss a e we e used. Ma e al. [
5
] p epa ed an
e o sensi i i y model based on spa ial and empo al ea u es o es ima e he obus ness o ideos o
di e en packe loss scena ios.
The pape by Søgaa d e al. [
6
] sugges ed a eg ession equa ion o ideo quali y es ima ion wi h
a PCC oscilla ing be ween 0.7 o 0.9, based on he ideo con en ype. The main bene i o he pape by
Loh and Bong [
7
] is he enhancemen o he SSIM index. They inco po a ed he idea o spa ial and
empo al scene cha ac e is ics in o he SSIM me hodology. They ob ained an imp o emen in p ecision
bu , on he o he hand, he compu a ional ime o hei me ic doubled compa ed o he “baseline”
pe o mance o SSIM. The successo o ideo codec H.264, namely H.265/HEVC (high-e iciency ideo
coding), bu wi hou ul a-high de ini ion (UHD) sequences included in he es ing da ase , was aken
in o accoun o c ea e a eg ession o p edic ion o he subjec i e sco e in [
8
,
9
] wi h an ob ained PCC
o 0.92.
Mus a a and Hameed [
10
] had many es ing scena ios (packe loss a e, a ious bi a es and
scenes) bu wo ked only wi h low esolu ion o ideo encoding (H.264 codec). They applied a new
me hodology o nume ous machine lea ning applica ions (neu al ne wo k, naï e Bayes, o decision
ee) and ob ained a classi ica ion a e anging om 0.86 o 0.88. The a icle by Akh a e al. [
11
]
summa izes he ecen inno a ions in his esea ch opic and s udies se e al app oaches such as
linea and nonlinea associa ions be ween he QoE and QoS pa ame e s. Bampis and Bo ik [
12
] used
empo al and spa ial in o ma ion indexes as inpu s in o he model de eloping p ocess and gained a
high le el o model e aci y o abou 0.9 (PCC). The a icle by Gu e al. [
13
] deal wi h he no- e e ence
es ima ion model. They used inpu ec o elemen s such as con as , sha pness, and b igh ness.
They e i ied hei model on six ideo da abases and ob ained a classi ica ion a e oscilla ing be ween
0.73 and 0.9 acco ding o he pa icula es da abase. The bene i s and d awbacks o backp opaga ion
neu al ne wo k usage o ideo quali y es ima ion a e desc ibed in [
14
]. The au ho s pe o med
se e al es scena ios and pu o wa d he imp o emen schemes o a neu al ne wo k. They p epa ed
a selec ion o impo an ideo sequence cha ac e is ics and analyzed hei impac on quali y p edic ion.
The p oposed model ope a ed wi h a PCC o abou 0.91.
The au ho s o his p oposed a icle also con ibu ed o his esea ch. We designed a hyb id
me hod o IPTV quali y e alua ion based on a backp opaga ion neu al ne wo k [
2
]. Ou applica ion
akes con en ype, bi a e, packe loss a e, and esolu ion in o conside a ion as inpu elemen s,
ollowed by he popula ideo o ma s (H.264 and H.265/HEVC). A he ime o w i ing, ou model
can es ima e he end-use pe cep ion o pic u e quali y o bo h ideo p o iles concu en ly, which is
some hing ha none o he s a e-o - he-a models do.
All pape s desc ibed in his sec ion ied o de elop a model o sol ing pa e n ecogni ion
p oblems. The au ho s wan ed o ind egula i ies (pa e ns) in da a ob ained om subjec i e es ing
and quali a i e pa ame e s o ideo sequences. This s udy con ibu es o illing his gap. We chose
a sel -o ganizing map due o i s abili y o lea ning wi hou a supe ision model and clus e ing.
We wan ed o inspec he sui abili y o unsupe ised lea ning and clus e ing echniques o ideo
quali y p edic ion. As a ype o neu al ne wo k, a Kohonen map o e s gene aliza ion abili y o es ima e
he da a i has no ained on. A Kohonen map is easy o c ea e, and he aining p ocess does no
equi e deep knowledge abou machine lea ning.
Howe e , se ice p o ide s need o know i he o e ed ideo s eam is o su icien quali y o no .
P ecise in o ma ion on he MOS a ing does no play a majo ole in ne wo k adminis a o decisions
ega ding ne wo k se ings. Ou p oposed sys em has all he ad an ages o ou p e iously published
model, namely, almos eal- ime quali y es ima ion o bo h o he mos popula ideo codecs.
Symme y 2020,12, 1535 4 o 16
3. Me hodology
In ou las pape ela ed o his esea ch opic [
2
], we p oposed a dis o ed ideo da abase whe e
each o he es ideo sequences ob ained a sco e ex ac ed om subjec i e and objec i e e alua ion
p ocedu es. We selec ed he objec i e es ing me hodology SSIM o i s good ep esen a ion o human
quali y pe cep ion. Ano he bene i o he SSIM me ic is i s scale ange (0–1), which is easy o
no malize o success ul used in ne wo k aining. SSIM is a ull e e ence me ic, so i equi es o iginal
undis o ed ( e e ence) ideo sequences o simila i y calcula ion. The inal sco e anges om 0 ( o ally
di e en samples) o 1 ( wo samples ha a e exac ly alike). The ideo quali y in es iga ion p ocess
is a ec ed by h ee componen s. Luminance land con as ca e measu ed and compa ed, ollowed
by s uc u al compa ison s. The o e all index is a mul iplica i e combina ion o hese componen s,
as depic ed in Figu e 1. The simila i y measu e o wo sequences ( e e ence xand es sequence y) can
be exp essed as ollows [1,2]:
SIM(x,y)=[l(x,y)]α[c(x,y)]β[s(x,y)]γ, (1)
whe e exponen s
α
>0,
β
>0, and
γ
>0 measu e he weigh o each componen . The de aul se ing is
α=β=γ=1.
Symme y 2020, 12, x FOR PEER REVIEW 4 o 17
3. Me hodology
In ou las pape ela ed o his esea ch opic [2], we p oposed a dis o ed ideo da abase whe e
each o he es ideo sequences ob ained a sco e ex ac ed om subjec i e and objec i e e alua ion
p ocedu es. We selec ed he objec i e es ing me hodology SSIM o i s good ep esen a ion o
human quali y pe cep ion. Ano he bene i o he SSIM me ic is i s scale ange (0–1), which is easy
o no malize o success ul used in ne wo k aining. SSIM is a ull e e ence me ic, so i equi es
o iginal undis o ed ( e e ence) ideo sequences o simila i y calcula ion. The inal sco e anges
om 0 ( o ally di e en samples) o 1 ( wo samples ha a e exac ly alike). The ideo quali y
in es iga ion p ocess is a ec ed by h ee componen s. Luminance l and con as c a e measu ed and
compa ed, ollowed by s uc u al compa ison s. The o e all index is a mul iplica i e combina ion o
hese componen s, as depic ed in Figu e 1. The simila i y measu e o wo sequences ( e e ence x and
es sequence y) can be exp essed as ollows [1,2]:
𝑆𝐼𝑀(𝑥,𝑦)= 󰇟𝑙(𝑥,𝑦)󰇠󰇟𝑐(𝑥,𝑦)󰇠󰇟𝑠(𝑥,𝑦)󰇠, (1)
whe e exponen s α > 0, β > 0, and γ > 0 measu e he weigh o each componen . The de aul se ing is
α = β = γ = 1.
Figu e 1. SSIM me ic scheme.
The absolu e ca ego y a ing (ACR) is an assessmen me hod de eloped by he In e na ional
Telecommunica ion Union (ITU). The es sequences a e shown sepa a ely, i.e., one a a ime. Each
es sequence has o be e alua ed by eal obse e s. The esul s a e epo ed as a MOS alue ( he
a ing scale is shown in Table 1). Figu e 2 shows he whole p ocedu e. As can be seen, he o ing
limi is app oxima ely 10 s. The ACR me hod ep esen s he eal si ua ion be e because end-use s
canno compa e he ecei ed ideo s eam wi h he o iginal s eam made by he con en owne (e.g.,
a TV s a ion). The es en i onmen (ligh ing condi ions, iewing dis ance) wi h a 24” Dell P2415Q
UHD me he condi ions speci ied in he ecommenda ions [15]. The ecommended numbe o
obse e s is a leas 15, bu we had 60 iewe s who pa icipa ed in his expe imen (be ween 18 and
40 yea s old), wi h men p edomina ing by 39:21. Viewe s had a sho b eak e e y 30 min, and he
maximum es session du a ion was 1.5 h [15].
Table 1. The a ing scale o mean opinion sco e (MOS) [15].
MOS Quali y Ra ing Impai men
5 Bes Impe cep ible
4 High Pe cep ible e o , no annoying
3 Medium Sligh ly annoying ( isible e o )
2 Low Annoying ( isible e o )
1 Poo Ve y annoying ( isible e o )
Figu e 1. SSIM me ic scheme.
The absolu e ca ego y a ing (ACR) is an assessmen me hod de eloped by he In e na ional
Telecommunica ion Union (ITU). The es sequences a e shown sepa a ely, i.e., one a a ime. Each es
sequence has o be e alua ed by eal obse e s. The esul s a e epo ed as a MOS alue ( he a ing
scale is shown in Table 1). Figu e 2shows he whole p ocedu e. As can be seen, he o ing limi is
app oxima ely 10 s. The ACR me hod ep esen s he eal si ua ion be e because end-use s canno
compa e he ecei ed ideo s eam wi h he o iginal s eam made by he con en owne (e.g., a TV
s a ion). The es en i onmen (ligh ing condi ions, iewing dis ance) wi h a 24” Dell P2415Q UHD
me he condi ions speci ied in he ecommenda ions [
15
]. The ecommended numbe o obse e s is
a leas 15, bu we had 60 iewe s who pa icipa ed in his expe imen (be ween 18 and 40 yea s old),
wi h men p edomina ing by 39:21. Viewe s had a sho b eak e e y 30 min, and he maximum es
session du a ion was 1.5 h [15].
Table 1. The a ing scale o mean opinion sco e (MOS) [15].
MOS Quali y Ra ing Impai men
5 Bes Impe cep ible
4 High Pe cep ible e o , no annoying
3 Medium Sligh ly annoying ( isible e o )
2 Low Annoying ( isible e o )
1 Poo Ve y annoying ( isible e o )
Symme y 2020,12, 1535 5 o 16
Symme y 2020, 12, x FOR PEER REVIEW 5 o 17
Figu e 2. Absolu e ca ego y a ing (ACR) es ing p ocedu e [15].
The ideo con en ype can be de ined by he empo al (objec s mo ion) and spa ial (luminance)
in o ma ion (TI/SI). The ecommenda ion [16] ecognizes se e al ypes o ideo con en ega ding o
hese wo indexes. Ou model hen es ima es he subjec i e sco e (MOS) by ex ac ion o in o ma ion
om he si ua ion in he ne wo k (packe loss a e), ideo encoding pa ame e s (i.e., bi a e, codec
ype) and he con en ype.
The Shanghai Jiao Tong Uni e si y esea ch g oup made a ailable hei da abase o
uncomp essed UHD ideo sequences ha con ain a ious ideo con en ypes [17]. These ideo
sequences ha e a du a ion o 10 s wi h a ame a e o 30/s.
Fi s , selec ed ideo sequences we e downloaded in UHD esolu ion (3840 × 2160), 4:2:0 colo
sampling wi h 8-bi colo dep h (YUV o ma ). These se ings ep esen s anda d TV b oadcas ing
p o ile. Secondly, all ideo sequences we e encoded o ideo p o iles, namely H.264/AVC and
H.265/HEVC, by using he FFmpeg ool e sion 4.2 (includes x264 and x265 encode s), which allows
o modi ying bi a es and esolu ions. A s eaming p ocess was pe o med by a combina ion o
FFmpeg (as a s eaming se e ) and VLC Playe so wa e ( e sion 3.0.6— ecei ing side). We
cap u ed and sa ed he ideo s eam ansmi ed ia he local ne wo k in e ace using VLC Playe .
Du ing he s eaming p ocedu e, we ini ially se he packe loss o 0.1% ( he applica ion d ops
andomly selec ed packe s) a he local in e ace. Then we epea ed his s ep o packe loss in
inc emen s o 0.2%, 0.3%, 0.5%, 0.75%, and 1%. The s eaming p ocess simula ed he RTP/UDP/IP
con igu a ion (FFmpeg: -c copy - mpeg s udp://127.0.0.1:1234) wi h payload encapsula ion in MPEG-
TS ( anspo s eam) o ma ; hus, we ully adop ed he p inciples o IPTV s eam anspo a ion
o e he IP ne wo k [17]. The o al numbe o dis o ed ideo sequences was 432. The whole p ocess
o making he da ase and e alua ion pe o mance is depic ed in Figu e 3. The desc ip i e
cha ac e is ics o he chosen scenes a e displayed in Figu e 4. Fo mo e de ailed in o ma ion on he
p epa a ion o he es ing ideo sequences, please see ou o he pape [2].
Figu e 3. Da ase making p ocedu e.
Digi al ele ision e es ial b oadcas ing has a heo e ical bandwid h limi a ion o abou 31
Mbps pe single adio equency channel (64—Quad a u e ampli ude modula ion and 8 MHz wide
channel). As a esul o his es ic ion, he bi a e o one TV signal can oscilla e be ween 5 Mbps
(s anda d quali y), 10 Mbps (FullHD), and 15 Mbps (UHD o p emium quali y). Typically, se e al
TV signals a e b oadcas ed ia one adio channel (mul iplexing).
Figu e 2. Absolu e ca ego y a ing (ACR) es ing p ocedu e [15].
The ideo con en ype can be de ined by he empo al (objec s mo ion) and spa ial (luminance)
in o ma ion (TI/SI). The ecommenda ion [
16
] ecognizes se e al ypes o ideo con en ega ding o
hese wo indexes. Ou model hen es ima es he subjec i e sco e (MOS) by ex ac ion o in o ma ion
om he si ua ion in he ne wo k (packe loss a e), ideo encoding pa ame e s (i.e., bi a e, codec ype)
and he con en ype.
The Shanghai Jiao Tong Uni e si y esea ch g oup made a ailable hei da abase o uncomp essed
UHD ideo sequences ha con ain a ious ideo con en ypes [
17
]. These ideo sequences ha e a
du a ion o 10 s wi h a ame a e o 30/s.
Fi s , selec ed ideo sequences we e downloaded in UHD esolu ion (3840
×
2160), 4:2:0 colo
sampling wi h 8-bi colo dep h (YUV o ma ). These se ings ep esen s anda d TV b oadcas ing
p o ile. Secondly, all ideo sequences we e encoded o ideo p o iles, namely H.264/AVC and
H.265/HEVC, by using he FFmpeg ool e sion 4.2 (includes x264 and x265 encode s), which allows
o modi ying bi a es and esolu ions. A s eaming p ocess was pe o med by a combina ion o
FFmpeg (as a s eaming se e ) and VLC Playe so wa e ( e sion 3.0.6— ecei ing side). We cap u ed
and sa ed he ideo s eam ansmi ed ia he local ne wo k in e ace using VLC Playe . Du ing
he s eaming p ocedu e, we ini ially se he packe loss o 0.1% ( he applica ion d ops andomly
selec ed packe s) a he local in e ace. Then we epea ed his s ep o packe loss in inc emen s o 0.2%,
0.3%, 0.5%, 0.75%, and 1%. The s eaming p ocess simula ed he RTP/UDP/IP con igu a ion (FFmpeg:
-c copy - mpeg s udp://127.0.0.1:1234) wi h payload encapsula ion in MPEG-TS ( anspo s eam)
o ma ; hus, we ully adop ed he p inciples o IPTV s eam anspo a ion o e he IP ne wo k [
17
].
The o al numbe o dis o ed ideo sequences was 432. The whole p ocess o making he da ase and
e alua ion pe o mance is depic ed in Figu e 3. The desc ip i e cha ac e is ics o he chosen scenes a e
displayed in Figu e 4. Fo mo e de ailed in o ma ion on he p epa a ion o he es ing ideo sequences,
please see ou o he pape [2].
Symme y 2020, 12, x FOR PEER REVIEW 5 o 17
Figu e 2. Absolu e ca ego y a ing (ACR) es ing p ocedu e [15].
The ideo con en ype can be de ined by he empo al (objec s mo ion) and spa ial (luminance)
in o ma ion (TI/SI). The ecommenda ion [16] ecognizes se e al ypes o ideo con en ega ding o
hese wo indexes. Ou model hen es ima es he subjec i e sco e (MOS) by ex ac ion o in o ma ion
om he si ua ion in he ne wo k (packe loss a e), ideo encoding pa ame e s (i.e., bi a e, codec
ype) and he con en ype.
The Shanghai Jiao Tong Uni e si y esea ch g oup made a ailable hei da abase o
uncomp essed UHD ideo sequences ha con ain a ious ideo con en ypes [17]. These ideo
sequences ha e a du a ion o 10 s wi h a ame a e o 30/s.
Fi s , selec ed ideo sequences we e downloaded in UHD esolu ion (3840 × 2160), 4:2:0 colo
sampling wi h 8-bi colo dep h (YUV o ma ). These se ings ep esen s anda d TV b oadcas ing
p o ile. Secondly, all ideo sequences we e encoded o ideo p o iles, namely H.264/AVC and
H.265/HEVC, by using he FFmpeg ool e sion 4.2 (includes x264 and x265 encode s), which allows
o modi ying bi a es and esolu ions. A s eaming p ocess was pe o med by a combina ion o
FFmpeg (as a s eaming se e ) and VLC Playe so wa e ( e sion 3.0.6— ecei ing side). We
cap u ed and sa ed he ideo s eam ansmi ed ia he local ne wo k in e ace using VLC Playe .
Du ing he s eaming p ocedu e, we ini ially se he packe loss o 0.1% ( he applica ion d ops
andomly selec ed packe s) a he local in e ace. Then we epea ed his s ep o packe loss in
inc emen s o 0.2%, 0.3%, 0.5%, 0.75%, and 1%. The s eaming p ocess simula ed he RTP/UDP/IP
con igu a ion (FFmpeg: -c copy - mpeg s udp://127.0.0.1:1234) wi h payload encapsula ion in MPEG-
TS ( anspo s eam) o ma ; hus, we ully adop ed he p inciples o IPTV s eam anspo a ion
o e he IP ne wo k [17]. The o al numbe o dis o ed ideo sequences was 432. The whole p ocess
o making he da ase and e alua ion pe o mance is depic ed in Figu e 3. The desc ip i e
cha ac e is ics o he chosen scenes a e displayed in Figu e 4. Fo mo e de ailed in o ma ion on he
p epa a ion o he es ing ideo sequences, please see ou o he pape [2].
Figu e 3. Da ase making p ocedu e.
Digi al ele ision e es ial b oadcas ing has a heo e ical bandwid h limi a ion o abou 31
Mbps pe single adio equency channel (64—Quad a u e ampli ude modula ion and 8 MHz wide
channel). As a esul o his es ic ion, he bi a e o one TV signal can oscilla e be ween 5 Mbps
(s anda d quali y), 10 Mbps (FullHD), and 15 Mbps (UHD o p emium quali y). Typically, se e al
TV signals a e b oadcas ed ia one adio channel (mul iplexing).
Figu e 3. Da ase making p ocedu e.
Digi al ele ision e es ial b oadcas ing has a heo e ical bandwid h limi a ion o abou 31 Mbps
pe single adio equency channel (64—Quad a u e ampli ude modula ion and 8 MHz wide channel).
As a esul o his es ic ion, he bi a e o one TV signal can oscilla e be ween 5 Mbps (s anda d
quali y), 10 Mbps (FullHD), and 15 Mbps (UHD o p emium quali y). Typically, se e al TV signals a e
b oadcas ed ia one adio channel (mul iplexing).

Symme y 2020,12, 1535 6 o 16
Symme y 2020, 12, x FOR PEER REVIEW 6 o 17
(a) Wood (b) Camp i e Pa y
(c) Cons uc ion Field (d) Runne s
Figu e 4. Tes sequences [2,17] (clockwise): (a) a high-mo ion scene ( as came a o a ion); (b) nigh
scene (people si ing nex o a i e); (c) low-mo ion scene (mo ing o bulldoze , s a ic backg ound);
(d) Shanghai ma a hon (s a ic shoo ing).
Ou p e iously men ioned pape [2] inspec ed many ideo sequence ea u es wi h po en ial o
be a pa o he inpu da ase . The mo ion cha ac e is ics o a ideo ha e an impac on how codec can
mask missing da a du ing he ideo econs uc ion. In a monoch oma ic s a ic pa o he ame (e.g.,
sky o g ass), i is easy o calcula e he missing blocks o he decoding o he ideo ame. Howe e ,
in an ac ion mo ie, scenes change e y o en. As a esul , missing da a cause isible comp ession
a i ac s. Codec wi h a high comp ession a io (in ou case, H.265) is mo e sensi i e o da a losses
because each block ca ies mo e in o ma ion han he codec wi h a lowe comp ession a io (H.264).
The bi a e de ines how much isual in o ma ion is encoded, ypically pe second. Highe esolu ion
equi es a be e bi a e han a low- esolu ion ideo does, bu a e y high bi a e (e.g., 15 Mbps) is
coun e p oduc i e i low esolu ion is selec ed ( edundan in o ma ion ha canno imp o e he
isual quali y) [2,18].
A comple e lis o he chosen pa ame e s, as well as he p ojec ed model ou pu , is gi en in Table
2. Elemen s o he inpu ec o a e in bold, while MOS (in i alics) cha ac e izes cus ome subjec i e
opinion. As can be seen in Table 2, esolu ion and codec ype we e no included in he se o inpu s.
We used a ea u e selec ion echnique ha p o ed ha he bold ace pa ame e s ha e a se ious e ec
on clus e making. One o he well-known s a is ical me hods o inpu ec o dimensionali y
educ ion ( ea u e selec ions) is called p incipal componen analysis (PCA). PCA is an algo i hm
allowing us o educe he se o a iables in a way ha analyses he po en ial co ela ion be ween
hem. As a esul o his me hod, a se o a iables called p incipal componen s is p o ided, whe e
none o hese a iables a e c oss-co ela ed [19–21]. We e i ied he esul s o his s anda d p ocedu e
by U-ma ix isualiza ion (an addi ional ad an age o Kohonen maps) o ind he minimum se o
inpu ec o elemen s. In o de o educe edundancy in he da a space (po en ial co ela ion), SOM
can be used by conside ing he so-called componen (o weigh ) planes. We es ed all inpu s lis ed in
Table 2; acco ding o he gene a ed planes isualiza ion, we selec ed ou inpu s wi hou any isual
mu ual co ela ion. We can decla e ha unselec ed inpu aspi an s we e app oxima ed by he
a iables p esen ed in bold in Table 2.
Figu e 4.
Tes sequences [
2
,
17
] (clockwise): (
a
) a high-mo ion scene ( as came a o a ion); (
b
) nigh
scene (people si ing nex o a i e); (
c
) low-mo ion scene (mo ing o bulldoze , s a ic backg ound);
(d) Shanghai ma a hon (s a ic shoo ing).
Ou p e iously men ioned pape [
2
] inspec ed many ideo sequence ea u es wi h po en ial o be
a pa o he inpu da ase . The mo ion cha ac e is ics o a ideo ha e an impac on how codec can mask
missing da a du ing he ideo econs uc ion. In a monoch oma ic s a ic pa o he ame (e.g., sky
o g ass), i is easy o calcula e he missing blocks o he decoding o he ideo ame. Howe e ,
in an ac ion mo ie, scenes change e y o en. As a esul , missing da a cause isible comp ession
a i ac s. Codec wi h a high comp ession a io (in ou case, H.265) is mo e sensi i e o da a losses
because each block ca ies mo e in o ma ion han he codec wi h a lowe comp ession a io (H.264).
The bi a e de ines how much isual in o ma ion is encoded, ypically pe second. Highe esolu ion
equi es a be e bi a e han a low- esolu ion ideo does, bu a e y high bi a e (e.g., 15 Mbps) is
coun e p oduc i e i low esolu ion is selec ed ( edundan in o ma ion ha canno imp o e he isual
quali y) [2,18].
A comple e lis o he chosen pa ame e s, as well as he p ojec ed model ou pu , is gi en in Table 2.
Elemen s o he inpu ec o a e in bold, while MOS (in i alics) cha ac e izes cus ome subjec i e
opinion. As can be seen in Table 2, esolu ion and codec ype we e no included in he se o inpu s.
We used a ea u e selec ion echnique ha p o ed ha he bold ace pa ame e s ha e a se ious e ec on
clus e making. One o he well-known s a is ical me hods o inpu ec o dimensionali y educ ion
( ea u e selec ions) is called p incipal componen analysis (PCA). PCA is an algo i hm allowing us
o educe he se o a iables in a way ha analyses he po en ial co ela ion be ween hem. As a
esul o his me hod, a se o a iables called p incipal componen s is p o ided, whe e none o hese
a iables a e c oss-co ela ed [
19
–
21
]. We e i ied he esul s o his s anda d p ocedu e by U-ma ix
isualiza ion (an addi ional ad an age o Kohonen maps) o ind he minimum se o inpu ec o
elemen s. In o de o educe edundancy in he da a space (po en ial co ela ion), SOM can be used by
conside ing he so-called componen (o weigh ) planes. We es ed all inpu s lis ed in Table 2; acco ding
o he gene a ed planes isualiza ion, we selec ed ou inpu s wi hou any isual mu ual co ela ion.
We can decla e ha unselec ed inpu aspi an s we e app oxima ed by he a iables p esen ed in bold
in Table 2.
Symme y 2020,12, 1535 7 o 16
Table 2. Lis o da ase a iables.
Pa ame e Desc ip ion
Codec H.264/AVC, H.265/HEVC
Bi a e (Mbps) 5, 10, 15
Packe loss a e (%) 0.1, 0.2, 0.3, 0.5, 0.75, 1
Resolu ion HD, FullHD, UHD a
Full e e ence me ic SSIM
Con en ype S a ic scene, scene wi h signi ican mo ion, nigh , and spo s scene
ACR MOS alue
a
HD =high de ini ion (1280
×
720). FullHD and Ul aHD deno e esolu ions o 1920
×
1080 o 3840
×
2160, espec i ely.
The inpu ec o xconsis s o he pa ame e s se o segmen n, as de ined in he example below:
xn=













Bi a e
Packe loss
SSIM
Ca ego y o scene














=













5
0.1
0.969
1














. (2)
The ca ego y o he scene pa ame e is 1 o s a ic, 2 o a nigh scene, 3 o spo , and 4 o a highly
dynamic scene. Nis he numbe o segmen s in a session. In his case, he inpu ma ix composed o
he ec o s o all he segmen s is as ollows:
I=[x1,. . . xn,. . . xN]. (3)
3.1. Sel -O ganizing Map (Kohonen Map)
A Kohonen (o sel -o ganizing) map ies o ca ego ize inpu s based on hei simila i y in he inpu
space. This app oach di e s om lea ning wi h a eache due o i s weigh modi ica ion. A Kohonen
map uses compe i i e lea ning. Ou pu neu ons compe e among hemsel es. The neu on ha “wins”
his compe i ion ( he so-called “winne akes all” neu on) is ac i e and i s weigh ec o is upda ed
(and nodes in i s neighbo hood, oo).
SOM can be used o many classi ica ion p oblems because compe i i e lea ning makes clus e s
o neu ons (wi h simila cha ac e is ics), while each o he o med g oups can be conside ed as a
classi ica ion class.
A Kohonen map con ains only wo laye s, namely, inpu and ou pu (compe i i e) laye s. This map
allows us o ans o m a mul idimensional da ase in o a symme ical s uc u e o 2D nodes. In he
wo s -case scena io, his algo i hm c ea es as many classi ica ion classes as he compe i i e laye has
nodes. This ype o neu al ne wo k was c ea ed by Finnish p o esso Teu o Kohonen; he e o e, SOM
and Kohonen map a e in e changeable exp essions [
22
]. Unsupe ised lea ning does no eques a ge
ou pu s o be included in he da ase ; hus, no op imiza ion algo i hm ( o inding a minimum o he
cos unc ion) o linea co ela ion is p esen ed. Classi ica ion a e is he way o e alua e he accu acy
o he ained model. We measu ed he classi ica ion a e o di e en ne wo k opologies. The a e age
a e o success ac ions using SOM should be 85–90% o achie e excellen ou pu e aci y.
A success ully ained ne wo k can p edic he esponse om a se o inpu s in he o m o an
exac posi ion (2D g id xand ycoo dina es) o an exci ing node wi hin he ou pu laye . We can
decide i his node belongs o a pa icula clus e : clus e labelling depends on i s dominan con en
ep esen a ion—in ou case, he mos equen ly occu ing MOS sco e wi hin each clus e .
Kohonen Map Algo i hm
Fi s , he weigh ini ializa ion o each node was pe o med. The sample inpu ec o was chosen
om he aining da ase . Then he algo i hm ied o ind he neu on wi h he weigh ec o closes o
Symme y 2020,12, 1535 8 o 16
he inpu ec o . This neu on is called he bes ma ching uni (BMU), and i s calcula ion was based on
Euclidean dis ance (we also es ed Manha an ci y block dis ance, bu ob ained wo se esul s o all
es ed opologies), as ollows [23]:
dj= XN
i=0xi( )−wij( )2, (4)
whe e
dj
ep esen s indi idual elemen s o he inpu a iables ows, and
wij
ep esen s he weigh
be ween he i- h inpu and he j- h ou pu node. Then BMU is a node wi h he minimum Euclidean
dis ance:
dj∗=min(dj). (5)
Weigh adap a ion is gi en by his exp ession:
wij( +1)=wij( )+η( )h(j∗,j)xi( )−wij( ), (6)
whe e
η
is a lea ning a e and
h(j∗,j)
de ines weigh adap a ion wi hin a ce ain adius. Each i e a ion
makes he adius o neighbo hood unc ion dec ease. The basic heigh o neighbo hood unc ion h o
a Kohonen map is:
h(j∗,j)=






1, i d(i∗,i)≤ ( )
0, o he wise , (7)
whe e
d(i∗,i)
s ands o he dis ance be ween he winning neu on
i∗
and speci ic neu on
i
, and
is
de o ed o he adius. The opology may no be ci cula (g id) only. As depic ed in Figu e 5, symme ic
hexagonal opology is p e e ed, o ins ance, by MATLAB so wa e. We chose a ba ch algo i hm o
he aining p ocedu e. MATLAB c ea es one ba ch con aining all samples o aining da a. A he end
o he ba ch (i e a i e loop), weigh s a e upda ed and BMU is de e mined. This app oach is much
as e in compa ison o he sequen ial mode [24,25].
The las s ep o he aining is he alida ion o classi ie ou pu s o each es ed ne wo k opology.
U-ma ix is a use ul me hod o ne wo k ou pu s isualiza ion. The uni ied dis ance ma ix (U-ma ix)
shows he posi ion and size o he g ouped nodes, as shown in Figu e 6. A g aphic analysis is
o mula ed on he Euclidean me ic be ween he inpu s and nodes o a compe i i e laye (MATLAB
uses yellow o black scale). Yellow zones can be deno ed as clus e s and black (da k ed) pa s as
clus e bounda ies [
26
]. We conside he U-ma ix a use ul g aphical in e p e a ion o he esul s in case
someone a emp s o iden i y simila i ies in he inpu da ase and hey a e no amilia wi h machine
lea ning applica ions.
Symme y 2020, 12, x FOR PEER REVIEW 8 o 17
on Euclidean dis ance (we also es ed Manha an ci y block dis ance, bu ob ained wo se esul s o
all es ed opologies), as ollows [23]:
𝑑=

∑(𝑥(𝑡)−𝑤(𝑡))

 , (4)
whe e 𝑑 ep esen s indi idual elemen s o he inpu a iables ows, and 𝑤 ep esen s he weigh
be ween he i- h inpu and he j- h ou pu node. Then BMU is a node wi h he minimum Euclidean
dis ance: 𝑑∗=𝑚𝑖𝑛(𝑑). (5)
Weigh adap a ion is gi en by his exp ession:
𝑤(𝑡+1)=𝑤(𝑡)+𝜂(𝑡)ℎ(
𝑗
∗,
𝑗
)(𝑥(𝑡)−𝑤(𝑡)), (6)
whe e 𝜂 is a lea ning a e and ℎ(𝑗∗,𝑗) de ines weigh adap a ion wi hin a ce ain adius. Each
i e a ion makes he adius o neighbo hood unc ion dec ease. The basic heigh o neighbo hood
unc ion h o a Kohonen map is: ℎ(
𝑗
∗,
𝑗
)=1, 𝑖𝑓 𝑑(𝑖∗,𝑖)≤𝑟(𝑡)
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 , (7)
whe e 𝑑(𝑖∗,𝑖) s ands o he dis ance be ween he winning neu on 𝑖∗ and speci ic neu on 𝑖, and 𝑟 is
de o ed o he adius. The opology may no be ci cula (g id) only. As depic ed in Figu e 5,
symme ic hexagonal opology is p e e ed, o ins ance, by MATLAB so wa e. We chose a ba ch
algo i hm o he aining p ocedu e. MATLAB c ea es one ba ch con aining all samples o aining
da a. A he end o he ba ch (i e a i e loop), weigh s a e upda ed and BMU is de e mined. This
app oach is much as e in compa ison o he sequen ial mode [24,25].
The las s ep o he aining is he alida ion o classi ie ou pu s o each es ed ne wo k
opology. U-ma ix is a use ul me hod o ne wo k ou pu s isualiza ion. The uni ied dis ance ma ix
(U-ma ix) shows he posi ion and size o he g ouped nodes, as shown in Figu e 6. A g aphic analysis
is o mula ed on he Euclidean me ic be ween he inpu s and nodes o a compe i i e laye (MATLAB
uses yellow o black scale). Yellow zones can be deno ed as clus e s and black (da k ed) pa s as
clus e bounda ies [26]. We conside he U-ma ix a use ul g aphical in e p e a ion o he esul s in
case someone a emp s o iden i y simila i ies in he inpu da ase and hey a e no amilia wi h
machine lea ning applica ions.
Figu e 5. A ea o weigh modi ica ion educing o e ime.
Figu e 5. A ea o weigh modi ica ion educing o e ime.
Symme y 2020,12, 1535 9 o 16
Symme y 2020, 12, x FOR PEER REVIEW 9 o 17
Figu e 6. Clus e iden i ica ion by U-ma ix.
F om he ga he ed da a [2] ha is shown in Table 3, we se a SSIM in e al ange ela ed o each
MOS a ing. The whole p ocedu e o da a p epa a ion and modelling is shown in Figu e 7. I he
subjec i e a ing is wo se han 3 on he MOS scale, he deli e ed ideo s eam is o poo quali y, wi h
conside able a i ac s appea ing in he image (blu ing, inging a i ac s, e c.)
The ob ained esul s show ha a MOS alue o 4 o highe migh co ela e o a SSIM sco e o
0.98; a MOS a ing be ween 3 and 4 belongs o he in e al 0.95–0.979, and he es is lowe han 3 on
he MOS scale.
Figu e 7. The diag am o sel -o ganizing map (SOM) classi ie modelling.
Figu e 6. Clus e iden i ica ion by U-ma ix.
F om he ga he ed da a [
2
] ha is shown in Table 3, we se a SSIM in e al ange ela ed o each
MOS a ing. The whole p ocedu e o da a p epa a ion and modelling is shown in Figu e 7. I he
subjec i e a ing is wo se han 3 on he MOS scale, he deli e ed ideo s eam is o poo quali y,
wi h conside able a i ac s appea ing in he image (blu ing, inging a i ac s, e c.)
Table 3.
Ex ac ed s uc u al simila i y index (SSIM) sco es ela ed o he MOS scale ob ained om
subjec i e es ing. Da a a e aken om ou p e ious pape [2].
Video Sequence MOS In e als (ACR) H.264 (SSIM) H.265 (SSIM)
Camp i e pa y
≥4 1–0.95 1–0.98
≥3<4 0.949–0.92 0.979–0.95
≥2<3 0.919–0.78 0.949–0.885
Cons uc ion ield
≥4 1–0.98 1–0.98
≥3<4 0.979–0.94 0.979–0.96
≥2<3 0.939–0.865 0.959–0.915
Runne s
≥4 1–0.98 N/Aa
≥3<4 0.979–0.94 N/A
≥2<3 0.939–0.86 0.95–0.91
Wood
≥4 1–0.96 N/A
≥3<4 0.959–0.89 N/A
≥2<3 0.889–0.615 0.93–0.725
aN/A=No Applicable. MOS alue was associa ed wi h none o he compu ed SSIM indexes.
Symme y 2020, 12, x FOR PEER REVIEW 9 o 17
Figu e 6. Clus e iden i ica ion by U-ma ix.
F om he ga he ed da a [2] ha is shown in Table 3, we se a SSIM in e al ange ela ed o each
MOS a ing. The whole p ocedu e o da a p epa a ion and modelling is shown in Figu e 7. I he
subjec i e a ing is wo se han 3 on he MOS scale, he deli e ed ideo s eam is o poo quali y, wi h
conside able a i ac s appea ing in he image (blu ing, inging a i ac s, e c.)
The ob ained esul s show ha a MOS alue o 4 o highe migh co ela e o a SSIM sco e o
0.98; a MOS a ing be ween 3 and 4 belongs o he in e al 0.95–0.979, and he es is lowe han 3 on
he MOS scale.
Figu e 7. The diag am o sel -o ganizing map (SOM) classi ie modelling.
Figu e 7. The diag am o sel -o ganizing map (SOM) classi ie modelling.
Symme y 2020,12, 1535 16 o 16
12.
Bampis, C.G.; Li, Z.; Bo ik, A.C. Spa io empo al Fea u e In eg a ion and Model Fusion o Full Re e ence
Video Quali y Assessmen . IEEE T ans. Ci cui s Sys . Video Technol. 2019,29, 2256–2270. [C ossRe ]
13.
Gu, K.; Tao, D.; Qiao, J.; Lin, W. Lea ning a No-Re e ence Quali y Assessmen Model o Enhanced Images
wi h Big Da a. IEEE T ans. Neu al Ne w. Lea n. Sys . 2018,29, 1301–1313. [C ossRe ]
14. Yuana, Y.; Wang, C. IPTV ideo quali y assessmen model based on neu al ne wo k. J. Vis. Commun. Image
Rep esen . 2019,64, 102629. [C ossRe ]
15.
In e na ional Telecommunica ions Union. ITU-T P.913. Me hods o he Subjec i e Assessmen o Video
Quali y. Audio Quali y and Audio isual Quali y o In e ne Video and Dis ibu ion Quali y Tele ision in Any
En i onmen . 2016. A ailable online: h ps://www.i u.in / ec/T-REC-P.913/en (accessed on 12 Augus 2020).
16.
In e na ional Telecommunica ions Union. ITU-T P.910. Subjec i e Video Quali y Assessmen Me hods o
Mul imedia Applica ions. 2008. A ailable online: h ps://www.i u.in / ec/T-REC-P.910/en (accessed on
12 Augus 2020).
17.
Song, L.; Tang, X.; Zhang, W.; Yang, X.; Xia, P. The SJTU 4K ideo sequence da ase . In P oceedings o he 5 h
In e na ional Wo kshop on Quali y o Mul imedia Expe ience (QoMEX), Klagen u , Aus ia, 3–5 July 2013.
18.
Paudyal, P.; Ba is i, F.; Ca li, M. Impac o ideo con en and ansmission impai men s on quali y o
expe ience. Mul imed. Tools Appl. 2016,75, 16461–16485. [C ossRe ]
19.
Cao, L.J.; Chua, K.S.; Chong, W.K.; Lee, H.P.; Gu, Q.M. A compa ison o PCA, KPCA and ICA o
dimensionali y educ ion in suppo ec o machine. Neu ocompu ing 2003,55, 321–336. [C ossRe ]
20.
Belkin, M.; Niyogi, P. Laplacian eigenmaps o dimensionali y educ ion and da a ep esen a ion.
Neu al Compu . 2003,15, 1373–1396. [C ossRe ]
21.
Mousas, C.; Newbu y, P.; Anagnos opoulos, C.N. E alua ing he co a iance ma ix cons ain s o da a-d i en
s a is ical human mo ion econs uc ion. In P oceedings o he 30 h Sp ing Con e ence on Compu e G aphics,
Smolenice, Slo akia, 28–30 May 2014; Associa ion o Compu ing Machine y: New Yo k, NY, USA; pp. 99–106.
22. Kohonen, T. The sel -o ganizing map. Neu ocompu ing 1998,21, 1–6. [C ossRe ]
23.
Lendel, V.; Panciko a, L.; Fala , L.; Ma cek, D. In elligen Modelling wi h Al e na i e App oach Applica ion
o Ad anced A i icial In elligence in o T a ic Managemen . Commun.-Sci. Le . Uni . Zilina
2017
,19, 36–42.
24.
Xiao ei, Q.; Lin, Y.; Kai, G.; Lin u, M.; Tao, F.; Shuangyin, R.; Meng, S. S a is ics-Enhanced Di ec Ba ch
G ow h Sel -O ganizing Mapping o E icien DoS A ack De ec ion. Neu ocompu ing 2016,182, 66–81.
25.
Weh ens, R.; K uisselb ink, J. Flexible Sel -O ganizing Maps in kohonen 3.0. J. S a . So w.
2018
,87, 1–18.
[C ossRe ]
26.
Yugenda , P.; Ra ishanka , K.R. Analysis o C owd Flow Pa ame e s Using A i icial Neu al Ne wo k.
T ansp. Telecommun. J. 2018,19, 335–345. [C ossRe ]
27.
Minihi, R.; AlSabbagh, H.M.; Al-Rizzo, H.; Al-Oma y, A. End-To-End Th oughpu o Vane wi h and wi hou
Cloud E ec . T ansp. Telecommun. J. 2019,20, 52–61. [C ossRe ]
28.
Suying, Y.; Guo, Y.; Chen, Y.; Xie, F.; Yu, C.; Liu, Y. Enabling QoE Lea ning and P edic ion o WebRTC Video
Communica ion in WiFi Ne wo ks. 2016. A ailable online: h p://eeweb.poly.edu/ acul y/yongliu/docs/
yishuai_icc17.pd (accessed on 2 Sep embe 2020).
29.
Peng, B.; Lei, J.; Fu, H.; Shao, L.; Huang, Q. A Recu si e Cons ained F amewo k o Unsupe ised Video
Ac ion Clus e ing. IEEE T ans. Ind. In o m. 2020,16, 555–565. [C ossRe ]
30.
Bha , M.; Thiesse, J.; Le Calle , P. A Case S udy o Machine Lea ning Classi ie s o Real-Time Adap i e
Resolu ion P edic ion in Video Coding. In P oceedings o he 2020 IEEE In e na ional Con e ence on
Mul imedia and Expo (ICME), London, UK, 6–10 July 2020; pp. 1–6.
31.
Anwa , M.S.; Wang, J.; Khan, W.; Ullah, A.; Ahmad, S.; Fei, Z. Subjec i e QoE o 360-Deg ee Vi ual Reali y
Videos and Machine Lea ning P edic ions. IEEE Access. 2020,8, 148084–148099. [C ossRe ]
32.
Khokha , M.; Ehlinge , T.; Ba aka , C. F om Ne wo k T a ic Measu emen s o QoE o In e ne Video.
In P oceedings o he IFIP Ne wo king Con e ence, Wa saw, Poland, 20–22 May 2019; pp. 1–9.
©
2020 by he au ho s. Licensee MDPI, Basel, Swi ze land. This a icle is an open access
a icle dis ibu ed unde he e ms and condi ions o he C ea i e Commons A ibu ion
(CC BY) license (h p://c ea i ecommons.o g/licenses/by/4.0/).