Linking Bitstream Information to QoE: A Study on Still Images Using HEVC Intra Coding

INFORMATION AND COMMUNICATION TECHNOLOGIES AND SERVICES VOLUME: 17 |NUMBER: 4 |2019 |DECEMBER Linking Bi s eam In o ma ion o QoE: A S udy on S ill Images Using HEVC In a Coding Tomas MIZDOS1, Ma cus BARKOWSKY 2, Mi osla UHRINA1, Pe e POCTA1 1Depa men o Mul imedia and In o ma ion-Communica ion Technologies, Facul y o Elec ical Enginee ing and In o ma ion Technology, Uni e si y o Zilina, Uni e zi na 1, 010 26 Zilina, Slo ak Republic 2Depa men o In e ac i e Sys ems and In e ne o Things, Facul y o Applied Compu e Science, Deggendo Ins i u e o Technology, Die e -Go li z-Pla z 1, 94469 Deggendo , Ge many [email p o ec ed], ma cus.ba ko[email p o ec ed], mi osla[email p o ec ed], pe e .po[email p o ec ed] DOI: 10.15598/aeee. 17i4.3625 Abs ac . The coding ools used in image and ideo encode s aim a high pe cep ual quali y o low bi- a es. Analyzing he esul s o he encode s in e ms o quan iza ion pa ame e , image pa i ioning, p edic- ion modes o esiduals may p o ide impo an insigh in o he link be ween hose ools and he human pe cep- ion. As a i s s ep, his con ibu ion analyzes he pos- sibili y o anscode e e ence images o h ee well- known image da abases, i.e. IRCCyN/IVC, LIVE and TID2013, om hei o iginal, olde o ma s o HEVC; hus c ea ing a homogeneous da abase o 327 HEVC encoded images accompanied wi h bi s eam pa ame e s and alues ob ained om objec i e and subjec i e as- sessmen s. Secondly, i analyzes some o he HEVC in a coding pa ame e s ega ding hei in luence on he image quali y by using machine lea ning, namely Suppo Vec o Machine - Reg ession. Keywo ds HEVC, HEVC in a coding pa ame e s, image quali y. 1. In oduc ion Since he in oduc ion o i s s anda dized ideo cod- ing algo i hm ITU-T H.261, he ideo coding expe s ha e imp o ed he pe o mance by hal ing he lowe bi a e a he same pe cep ual quali y e e y ew yea s. Mos o his gain is due o imp o ed algo i hms o in- a and in e ame p edic ion and u he e inemen o esidual coding. The esul s also show ha he hu- man isual sys em is as sa is ied wi h an HEVC bi - s eam as i was wi h H.261 o mo e han en imes he size. The ques ion o he QoE communi y is hen: Wha can we lea n om he in o ma ion educ ion p o- cess in he ideo encode o he analysis o he image and ideo quali y? This pape may se e as a i s s ep owa ds ha ques ion by c ea ing a da ase o HEVC in a coded images om olde s ill image coding s anda ds such as JPEG. This s ep p o ides us wi h su icien ly la ge g oup o sou ce images, oughly associa ed o sco e ob- ained om subjec i e es s. Then, some i s p ope - ies o he HEVC bi s eam a e analyzed owa ds iden- i ying he impo ance o he highly lexible pa i ion- ing p ocess which is one o he s eng hs o HEVC. While he mo i a ion in his endea o is di e en , he wo k is also closely ela ed o No-Re e ence qual- i y measu emen and also can se es as a base o de- elopmen o new No-Re e ence quali y measu emen s. Fo his eason, we p o ide a b ie o e iew o hem. In ecen yea s many app oaches o link bi s eam pa ame e s o QoE we e in oduced [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12] and [13]. The quali y p edic ion app oaches de eloped o H.264/AVC coded ideos, especially he bi s eam based ones, a e no applicable o HEVC coded ideos [2]. A ew HEVC bi s eam based quali y es ima ion models ha e been al eady p oposed in he li e a u e. The model p o- posed by Lei e al. in [3] is based on a linea e- g ession and ex ac s Quan iza ion Pa ame e A e age and Skip Coding Uni Pe cen om HEVC bi s eam o es ima e a ideo quali y p edic ed by PSNR. In [4], Anegekuh e al. de eloped he eg ession model, based on a mou ion amoun me ic (a me ic desc ibed in his pape de e mining ideo con en ype) and Quan- iza ion Pa ame e (QP), being able o es ima e ideo c 2019 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 436 INFORMATION AND COMMUNICATION TECHNOLOGIES AND SERVICES VOLUME: 17 |NUMBER: 4 |2019 |DECEMBER quali y p edic ed again by PSNR. An ex ended e - sion o his model aking in o accoun also a complex- i y o ideo sequences was p oposed in [5]. Simila ly as in [4] and [5], Anegekuh e al. p oposed he e- g ession model based on he QP and con en ype in his wo k cha ac e ized by he con en ype classi i- ca ion me ic, es ima ing quali y alues p edic ed by PSNR. Shahid e al. designed in [6] he model, based on a wo-laye eed o wa d a i icial neu al ne wo k and in ol ing 43 HEVC bi s eam pa ame e s, being able o es ima e a ideo quali y p edic ed by PSNR, VQM, VIF, and PVQM. In [7] Izumi e al. p oposed an objec- i e pe cep ual ideo-quali y-measu emen o HEVC. They in oduced wo pa ame ic NR me hods o es i- ma e a pe cep ual pic u e quali y o HEVC. In [8], He e al. p oposed No-Re e ence model which conside s bi s eam and display pa ame e s, o be mo e gene - alized o di e en e minal scena ios. The ollowing pa ame e s we e conside ed as inpu s o he model: Quan isa ion Pa ame e , mo ion ec o s, ideo com- plexi y and key- ame indica o and display pa ame- e s ( esolu ion, PPI, . . . ). The au ho s ied o make gene al model o bo h H.264/AVC and H.265/HEVC s anda ds. In [9] Fazliani e al. p oposed a nea eal- ime No-Re e ence ideo quali y assessmen me hod. They ained a ully connec ed neu al ne wo k wi h ea u es ex ac ed om bo h bi s eam and pixel do- mains along wi h hei espec i e subjec i e quali y sco es. In [10], Alizadeh e al. p esen ed a no el No-Re e ence Video Quali y Assessmen (NR-VQA) based on Con olu ional Neu al Ne wo k (CNN) o he HEVC. In [11], Ren e al. p esen ed a No-Re e ence quali y assessmen algo i hm o UHD HEVC encoded ideos. The algo i hm di ec ly ex ac s he speci ic ideo cha ac e is ics om he HEVC bi s eam and es ablishes he ma hema ical model be ween he ideo cha ac e is ics and he inal ideo quali y h ough a lin- ea eg ession. The algo i hm ex ac s h ee ideo ea- u es, i.e. quan iza ion pa ame e s o he measu emen comp ession damage, numbe o CTUs o di e en sizes (8×8 & 32×32), and numbe s o he SAO blocks. Huang e al. p oposed in [12] a No-Re e ence (NR) Video Quali y Assessmen (VQA) me hod o ideos dis o ed by he HEVC. The assessmen was pe o med wi hou an access o a bi s eam. The p oposed analy- sis was based on he ans o m coe icien s es ima ed by he decoded ideo pixels, which a e used o es ima e a le el o quan iza ion. In [13], Nawada e al. p e- sen ed and desc ibed mo e quali y indica o s ha can be used in a No-Re e ence QoE calcula ion, since some o hem de ec speci ic e o s. Such e o s a e di icul o include in a global QoE model bu a e impo an om he ope a ion poin o iew. In his pape , we i s map HEVC Quan iza ion Pa- ame e s (QPs) using h ee well-known image qual- i y da abases, i.e. he IRCCyN/IVC [14] da abase, he LIVE image da abase [15], and he TID2013 da abase [16], o MOS. As a esul o he mapping p ocess, a new me ged da ase con aining 327 HEVC encoded images accompanied wi h he co esponding bi s eams and second o de MOS alues is c ea ed. Second o de MOS means indica i e quali y alues de i ed om he MOSs included in he men ioned da ase s by linea alignmen . In he second s ep, he new da ase is used o analyze an impac o some HEVC in a coding pa ame e s, i.e. QPs, dis ibu- ion o Coding Uni (CU), P edic ion Uni (PU) and T ans o m Uni (TU), on image quali y expe ienced by he end use by deploying Suppo Vec o Machine - Reg ession (SVM). The emainde o he pape is o ganized as ollows. Sec ion 2. desc ibes he expe imen dealing wi h he mapping o he HEVC QPs o MOS. In Sec. 3. , he analysis o he selec ed HEVC in a coding pa am- e e s impac on image quali y using SVM is p esen ed. Sec ion 4. p o ides he inal conclusions and sugges s u u e wo k. 2. Mapping o HEVC Images o Subjec i e Quali y o JPEG Images 2.1. Desc ip ion o he Used Da ase s We would like o ind ou whe he he e exis s a ela- ionship be ween bi s eam pa ame e s o in a coded H.265/HEVC images and hei subjec i e quali y. Un o una ely, he e is no su icien ly la ge da ase o HEVC in a coded images anno a ed by subjec i e quali y es s. We ha e decided o euse olde anno a ed da ase s wi h simila ype o dis o ions like he HEVC coding p oduces. Sui able da ase s should be hose which con ain images wi h JPEG deg ada ion because i in oduces simila ypes o dis o ions as HEVC coding. In ou expe imen s, h ee well-known image quali y da abases, namely IRCCyN/IVC [14] da abase, LIVE image da abase [15], and TID2013 da abase [16], con aining non-deg aded and also JPEG-deg aded im- ages, besides o he deg ada ions, we e used. All o he abo e men ioned da ase s in ol e esul s om subjec i e es s as well. In he case o he IRC- CyN/IVC da abase, he Double S imulus Impai men Scale (DSIS) [17] me hod was used o ob ain quali y sco es. This me hod uses a i e-g ade impai men scale whe e one means e y annoying and i e impe cep ible di e ence be ween he deg aded and e e ence image. The Absolu e Ca ego y Ra ing (ACR) [18] me hod was employed o a subjec i e e alua ion in he case o he LIVE image da abase. The g ading scale was di- ided in o i e linea equal egions anging om low c 2019 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 437 INFORMATION AND COMMUNICATION TECHNOLOGIES AND SERVICES VOLUME: 17 |NUMBER: 4 |2019 |DECEMBER IVC/IRCYN LIVE TID 2013 HM e e ence so wa e QP: 0-51 Well known da a Objec i e measu eme s (PSNR, SSIM, VIF) Compa ison o objec i e alues min(VQMHEVC-VQMJPEG)2 Median o VQM alues 327 HEVC encoded images INLSA MOS scaling o 1-5 Objec i e measu eme s (PSNR, SSIM, VIF) New ILT-HEVC da ase 327 images wi h second o de MOS 1-5 Re e ence images (Non-deg aded) JPEG dis o ed images PSNR, SSIM, VIF alues SSIM alues Subjec i e sco es (IVC/IRCyN MOS 1-5, LIVE MOS 1-100, TID2013 0-9) HEVC dis o ed images PSNR, SSIM, VIF alues PSNR minial di e ence SSIM minial di e ence VIF minial di e ence Fig. 1: Basic schema o da a p epa a ion p ocess. o excellen quali y. The scale was hen linea ly con- e ed in o 1–100. A pai wise so ing me hodology de- sc ibed in [16] was used in he case o he TID2013 da abase o ob ain a isual quali y. The subjec i e es s we e conduc ed in i e coun ies, i.e. Finland, Uk aine, F ance, I aly, and USA. The MOS alues ob- ained by his me hodology a y om 0 o 9 whe e he la ge alues co espond o be e isual quali y. Mo e de ails, like esolu ion, numbe o he used e e - ence and deg aded images a e p esen ed in he Tab. 1. Tab. 1: Pa ame e s o he used image da ase s. Da ase IRCCyN/IVC LIVE TID2013 Name Numbe 10 29 25 o images Numbe 50 175 125 o JPEG dis o ed images Resolu ion 512×512 768×512 512×384 ( ypically) 2.2. Da a P epa a ion o Expe imen Applying machine lea ning algo i hms equi e a la ge amoun o inpu da a. Due o his ac , we ha e me ged he h ee abo e men ioned da ase s in o la ge one. Sou ce ( e e ence) images wi hou any dis o ions se e as a base o he new bigge da ase . Whole p ocess o da a p epa a ion and c ea ion o new da ase sui - able o machine lea ning is depic ed in Fig. 1. Fi s ly, a colou space o each e e ence image was con e ed om RGB o YUV420p. Secondly, all he e e ence images om he da ase s we e encoded by he HEVC comp ession s anda d using he HM e e ence so wa e ( e sion 16.20) [19]. All he encoding se ings we e kep de aul , excep Quan isa ion Pa ame e . The Quan i- sa ion Pa ame e (QP) has been con inuously changed om 1 o 51 du ing he encoding p ocess. As a esul , a new da ase con aining 3 264 HEVC encoded images accompanied wi h he co esponding bi s eams was c ea ed. This da ase in ol es 51 le els o comp ession deg ada ion o each e e ence image. To in es iga e he ela ionship be ween HEVC s eam pa ame e s and quali y o image, i is necessa y o ha e quali y assess- men s o he co esponding images a hand. The bes way o ge hem is o execu e subjec i e es s bu his is also a e y expensi e and ime consuming app oach. Fo ou pu pose, an app oxima e alue o subjec i e sco e is su icien jus o see whe he i is possible o map HEVC in a coding pa ame e s o image quali y. So, ins ead o assessing he quali y o all he images subjec i ely, an equi alen quali y o he deg aded im- ages om he o iginal da abases was sough in o de o assign hei subjec i e MOS alues o he HEVC in a- coded images. In o he wo ds, we mapped a ailable quali y sco es o he JPEG image o he coun e pa HEVC coded image. In o de o ind ou a ela ionship be ween he HEVC and JPEG deg ada ions, h ee di - e en objec i e measu es, namely Peak Signal o Noise Ra io (PSNR), S uc u al SIMila i y index, and Visual In o ma ion Fideli y (VIF), we e compu ed. Acco d- ing o [20], he e is a ela ionship be ween objec i e measu e alues and subjec i e sco e. Men ioned objec- i e measu emen s we e done on bo h he new HEVC da ase and JPEG deg aded images. Fo calcula ing he PSNR, SSIM, and VIF alues, a Video Quali y Measu emen Tool (VQMT) de eloped by Mul imedia Signal P ocessing G oup ( e sion 1.1) was used in [21]. c 2019 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 438 INFORMATION AND COMMUNICATION TECHNOLOGIES AND SERVICES VOLUME: 17 |NUMBER: 4 |2019 |DECEMBER Dispesion 4; (Selec ed QP 23) 23 20 24 PSNR SSIM VIF 0 5 10 15 20 25 (a) Dispesion 4; (Selec ed QP 48) 49 48 45 PSNR SSIM VIF 0 10 20 30 40 50 (b) Dispesion 3; (Selec ed QP 42) 43 42 40 PSNR SSIM VIF 0 10 20 30 40 50 (c) Dispesion 3; (Selec ed QP 42) 43 42 40 PSNR SSIM VIF 0 10 20 30 40 50 (d) Fig. 2: Dis ibu ions o he QP alues ob ained by he measu es o he dispe sion wo s cases. 2.3. Mapping HEVC Deg ada ion o JPEG Subjec i e Sco e Residues be ween he HEVC and JPEG me ics esul s we e compu ed acco ding o ollowing Eq. (1): HEV C_QP = min(((V QM(JP EGimage)+ −V QM(HEV Cimage))2),(1) whe e HEV C_QP is he HEVC image wi h he clos- es objec i e quali y o he JPEG image and V QM ={P SNR, SSIM, V IF }. Each esul om he HEVC da ase was compa ed o he esul s om he JPEG da ase . This compa ison was done o each objec i e me hod sepa a ely. Fo he HEVC encoded image wi h he QP alue whe e he objec i e measu es ma ched bes (minimum esidual), he MOS alue o he JPEG deg aded image was assigned. As h ee di e en measu es (PSNR, SSIM, VIF) we e used, some imes happened ha h ee di e en candida es o he single JPEG sco e we e deno ed. I was caused by di e en app oaches and sensi i i y o he used ob- jec i e measu emen s. The di e en candida es co e- spond o HEVC images wi h sligh ly di e en alues o QP. The e o e, we ha e calcula ed dispe sion be- ween images selec ed by each objec i e measu emen acco ding o he ollowing Eq. (2): QP _Dispe sionp s = max(selec edQP (V QM))+ −min(selec edQP (V QM)), (2) whe e QP _Dispe sionp s is a dispe sion o each im- age and selec edQP (V QM)is he QP o he images selec ed by he co ensponding VQM. A maximum dis- pe sion ob ained by he h ee objec i e me ics was 4 QP uni s and ha e appea ed only wice. A di e ence o 4 QP uni s is isually ha dly no iceable. I is wo h no ing ha he di e ence is a he small conside ing he ac ha PSNR, SSIM, and VIF a e based on qui e di e en app oaches. The dispe sion and equency o occu ence o di e en QP alues a e depic ed in Fig. 3. Dis ibu ion o QP dispe sion 69 188 58 10 2 01234 QP dispe sion 0 20 40 60 80 100 120 140 160 180 200 Numbe o occu ance Fig. 3: Size o QP dispe sion s numbe o occu ence. We also analysed dis ibu ions o all he objec i e measu es deployed in ou expe imen . Figu e 2 depic s ou wo s cases in e ms o he QP dispe sion. As you can see om his pic u e, he alues ob ained om c 2019 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 439 INFORMATION AND COMMUNICATION TECHNOLOGIES AND SERVICES VOLUME: 17 |NUMBER: 4 |2019 |DECEMBER he objec i e measu es sligh ly di e . This is caused by di e en app oaches used by he measu es and also by a con en o he es images. In o de o ge only one image associa ed o he sub- jec i e quali y, a median alue o QPs p o ided by di - e en objec i e measu emen s was chosen and he co - esponding HEVC image was agged wi h he MOS sco e om he o iginal JPEG da ase . Figu e 4 depic s a pe cen age o he cases when he alue ob ained by he measu e di e s om he selec ed median alue. I is ob ious ha he esul s ob ained by PSNR mos ly di e om he selec ed median alue. The eason can be ha PSNR is a a he simple measu e which does no conside p ope ies o human isual sys em. As he MOS sco es we e no di ec ly ob ained by sub- jec i e es ing bu by he alignmen using he objec i e measu es, we will e e o hem as a second o de MOS in his wo k. 45% 34% 16% PSNR SSIM VIF 0 5 10 15 20 25 30 35 40 45 Fig. 4: Pe cen age o he cases when he measu es p o ided he di e en QP alue om he selec ed median alue. 012345 MOS 25 30 35 40 45 50 Median QP IVC MOS o median QP A ion Ba ba Boa s Clown F ui House Isabe Lena Mand il Pimen Fig. 5: Rela ionship be ween he HEVC QP and JPEG MOS o he IRCCyN/IVC da abase wi h deno ed di e en con en o images. By he abo e speci ied app oach, we map he Quan- isa ion Pa ame e o he images o alues desc ibing hei quali y, i.e. JPEG MOS. As he eused olde image quali y da ase s ha e included di e en MOS scales, we ha e execu ed a mapping o QP o quali y sco e on each da ase sepa a ely. A ela ionship be- ween QP and JPEG MOS is depic ed in Fig. 5, Fig. 6 and Fig. 7. 0 10 20 30 40 50 60 70 80 90 100 MOS 0 10 20 30 40 50 60 Median QP LIVE MOS o median QP Ligh house Sailing2 Sailing3 S a ue Woman Womanha Building2 Coinsin oun ain Flowe sonih35 Bikes Buildings Caps house ligh house2 mona ch ocean pain edhouse pa o s plane apids sailing1 sailing4 st eam ca ni aldolls ceme y chu chandcapi ol dance s manfishing studen sculp u e Fig. 6: Rela ionship be ween he HEVC QP and JPEG MOS o he LIVE image da abase wi h deno ed di e en con- en o images. 0123456789 MOS 0 10 20 30 40 50 60 Median QP TID MOS o median QP I01 I02 I03 I04 I05 I06 I07 I08 I09 I10 I11 I12 I13 I14 I15 I16 I17 I18 I19 I20 I21 I22 I23 I24 I25 Fig. 7: Rela ionship be ween he HEVC QP and JPEG MOS o he TID2013 da abase wi h deno ed di e en con en o images. Figu e 5, Fig. 6 and Fig. 7 show a ypical beha iou o JPEG MOS e sus QP, showing ha he mapping may be conside ed easonable. Howe e , he h ee dis inc da ase s a e s ill consid- e ed small in e ms o machine lea ning needs. c 2019 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 440 INFORMATION AND COMMUNICATION TECHNOLOGIES AND SERVICES VOLUME: 17 |NUMBER: 4 |2019 |DECEMBER 2.4. Da ase s Me ging Due o he equi emen o la ge da ase o mo e eli- able esul s, he MOS alues coming om he h ee da ase s we e aligned and me ged. To me ge he da ase s, an I e a ed Nes ed Leas -Squa es Algo i hm (INLSA) p oposed in [22] was deployed. Inpu s o he algo i hm we e objec i e alues o SSIM and subjec i e JPEG MOSs om all he h ee da ase s. The LIVE da ase was used as he e e ence one when i comes o he me ging p ocess. I means ha e- maining da ase s we e aligned acco ding o he LIVE da abase. The eason why he LIVE as e e ence was chosen is ha he LIVE da ase is he bigges one and con ains a highe ange o subjec i e MOS. As he INLSA u ilizes he unc ional ela ionships be ween an objec i e quali y me ic (ex ac ed om he images) and he co esponding subjec i e MOS, SSIM was used in he me ging p ocess as he inpu objec i e quali y me ic. In ou case, he INLSA al- go i hm calcula ed weigh s and scaling pa ame e s on he basis o SSIM alues and hen scaled MOS linea ly o 1–5 scale. Fi s ly, he o iginal MOS alues we e lin- ea ly ans o med o he in e al [0, 1] in each da ase sepa a ely. I is so called dis o ion domain, whe e 0 ep esen s no-impai men and 1 ep esen s se e e im- pai men . Then he scaling p ocess ollows a ypical linea Eq. (3): scaledMOS =a·MOS0+b, (3) whe e amean gain, bshi and MOS0is MOS in he dis o ion domain. As i was al eady men ioned abo e, he LIVE da abase was used as he e e ence da ase . I means ha a= 1 and b= 0. Fo he IRCCyN/IVC and he TID2013, we ha e ob ained a= 0.7696 and b= 0.2160 and a= 1.0475 and b=−0.1168 by he INLSA espec i ely. As a esul o he me ging p ocess, he new da ase en i led ILT- HEVC con aining 327 HEVC encoded images accom- panied wi h he co esponding bi s eams and second o de MOS alues was made. 2.5. Possible Limi a ions o he App oach I is wo h no ing he e ha he p oposed app oach migh ha e been posi i ely/nega i ely in luenced by a couple o e ec s. The me ging p ocess o he da ase s can be one o hem. We ha e used he INLSA o me ge he alues o he MOS by using he objec i e alues bu an unde s anding o he ask by he obse e s can be, a he end, di e en . Secondly, di e en subjec i e es me hodologies we e deployed o ge he MOS sco es when i comes o he da ase s. The di e en me hod- ologies led o he MOS sco es wi h di e se meaning and in di e en scales. Thi dly, he mapping based on he objec i e alues can be seen as one o he p ospec- i e sou ces o noise in his con ex . Despi e he ac ha we ha e deployed he ull e e ence objec i e me h- ods, which a e conside ed a he eliable, hey a e s ill a om being pe ec . I is wo h o ei e a e he e ha we ha e used h ee di e en ull e e ence measu es and selec ed he inal quali y on he basis o he median alue. Finally, a di e en appea ance o he JPEG and HEVC dis o ions can also ha e some in luence in his case. I is wo h no ing ha he JPEG comp ession uses ixed size o blocks (8×8 pixel) unlike he HEVC, which uses adap i e size. Highe size o blocks can lead o blu ing impo an lines and can subsequen ly cause wo se subjec i e assessmen . 3. Analysis o he Selec ed HEVC In a Coding Pa ame e s Impac on Image Quali y by SVM-Reg ession 3.1. Expe imen Desc ip ion Fi s , he ILT-HEVC da ase con aining 327 HEVC encoded images accompanied wi h he co espond- ing bi s eams and second o de MOS alues was di- ided in o wo pa s by he 80/20 a io, ypically de- ployed when i comes o he machine lea ning ap- p oaches. The da ase di ision was andom while con- side ing some es ic ions o a oid a model o e i ing. Thi een di e en con en s co e ing all he quali y le - els, anging om 4 o 7, we e andomly selec ed o a alida ion subse . I means ha 68 images, ou o he 327 dis o ed images, we e selec ed o a alida ion and he emaining images we e used o a aining o SVM model. I should be men ioned ha he con en deployed in he aining da ase was no eplica ed in he alida ion da ase . This ac should a oid o e i - ing. The SVM was selec ed as a machine lea ning echnique o be used o his analysis. Fo a SVM aining, we ha e used a unc ion included in MAT- LAB, en i led “ i s m”, which ains a Suppo Vec- o Machine (SVM) eg ession model on a low- h ough mode a e-dimensional p edic o da a se . Rega ding a ke nel unc ion, as second o de polynomial unc- ion has led o he bes esul s, his unc ion was used in his analysis as he ke nel unc ion. The epsilon pa ame e was se o 0.75. O he SVM se ings e- main unchanged (de aul ). Inpu o he SVM we e he QP, he Coding, T ans o m, and P edic ion Uni Size (CU, TU, PU). The a ge alue was he second o de MOS. The used unc ion has e u ned a ull SVM c 2019 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 441 INFORMATION AND COMMUNICATION TECHNOLOGIES AND SERVICES VOLUME: 17 |NUMBER: 4 |2019 |DECEMBER eg ession model ained by he selec ed HEVC in a coding pa ame e s and he co esponding e e ence o quali y alues he second o de MOS. The aining p o- cess was epea ed 9 imes wi h di e en combina ions o he HEVC in a coding pa ame e s in o de o ind ou how di e en HEVC in a coding pa ame e s in- luence image quali y. Pea son Linea Co ela ion Co- e icien (PLCC), Spea man Rank-O de Co ela ion Coe icien (SROCC), and Roo Mean Squa e E o (RMSE) we e used as pe o mance indica o s in his analysis. 3.2. Expe imen al Resul s Figu e 8, Fig. 9 and Fig. 10 compa e he second o de MOS alues wi h he p edic ions p o ided by he SVM eg ession model o he selec ed combina- ions o he HEVC in a coding pa ame e s in ol ed in he SVM-based eg ession p ocess. I can be obse ed om Fig. 8 ha he co ela ion esul s a e a he good when all he selec ed HEVC in a coding pa ame e s a e used. Mo eo e , he epo ed RMSE alue is also a he good. 1 1.5 2 2.5 3 3.5 4 4.5 5 Real MOS 1 1.5 2 2.5 3 3.5 4 4.5 5 SVR p edic ed MOS Real MOS o SVR p edic ed MOS (Polynomial ke nel q=2) PLCC = 0.9029 SROCC = 0.9226 RMSE = 0.4167 Fig. 8: Co ela ion be ween he second o de MOS alues and p edic ions p o ided by he SVM eg ession model o all he selec ed HEVC in a coding pa ame e s. The esul s o all he in es iga ed combina ions a e summa ized in Tab. 2. As i can be clea ly seen om his able, he bes esul s in e ms o RMSE we e achie ed o he combina ion in ol ing he QP, CU, and PU, ollowed by he QP, CU, and TU combina ion as well as he QP and CU combina ion. I is wo h no - ing he e ha he epo ed RMSE alue is e en smalle han ha ob ained o he combina ion in ol ing all he pa ame e s. Figu e 9 depic s he co ela ion be- ween he second o de MOS alues and p edic ions p o ided by he SVM eg ession model o he bes pe o ming pa ame e s combina ion in e ms o RMSE. When i comes o he combina ion in ol ing he CU, TU and PU pa ame e s and a numbe o he suppo Tab. 2: Resul s ob ained o he di e en combina ions o he in es iga ed HEVC in a coding pa ame e s. QP CU TU PU Pe o mance indica o s X PLCC = 0.90 SROCC = 0.92 RMSE = 0.45 Numbe o SVs = 28 X X PLCC = 0.92 SROCC = 0.93 RMSE = 0.37 Numbe o SVs = 26 X X PLCC = 0.92 SROCC = 0.93 RMSE = 0.38 Numbe o SVs = 26 X X PLCC = 0.90 SROCC = 0.93 RMSE = 0.38 Numbe o SVs = 24 X X X PLCC = 0.92 SROCC = 0.93 RMSE = 0.37 Numbe o SVs = 26 X X X PLCC = 0.92 SROCC = 0.93 RMSE = 0.37 Numbe o SVs = 24 X X X PLCC = 0.91 SROCC = 0.93 RMSE = 0.38 Numbe o SVs = 24 X X X PLCC = 0.85 SROCC = 0.86 RMSE = 0.52 Numbe o SVs = 48 X X X X PLCC = 0.90 SROCC = 0.92 RMSE = 0.42 Numbe o SVs = 27 1 1.5 2 2.5 3 3.5 4 4.5 5 Real MOS 1 1.5 2 2.5 3 3.5 4 4.5 5 SVR p edic ed MOS Real MOS o SVR p edic ed MOS (Polynomial ke nel q=2) Pa ame e s wi hou TU size PLCC = 0.9152 SROCC = 0.9284 RMSE = 0.3688 Fig. 9: Co ela ion be ween he second o de MOS alues and p edic ions p o ided by he SVM eg ession model o he QP, CU and PU dis ibu ions. ec o s, he numbe doubled in compa ison o he o he in es iga ed combina ions. We we e in e es ed in he dependency o he esul - ing pe o mance on he inpu alues. E en only by us- c 2019 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 442 INFORMATION AND COMMUNICATION TECHNOLOGIES AND SERVICES VOLUME: 17 |NUMBER: 4 |2019 |DECEMBER ing QP, we disco e ed ha he anking o de was s ill easonable as i can be seen in Fig. 10. I should be no ed he e ha he esul s ob ained in his s udy we e achie ed by a a he small da ase . So, a s a is ical signi icance o he esul s is a he limi ed. 1 1.5 2 2.5 3 3.5 4 4.5 5 Real MOS 1 1.5 2 2.5 3 3.5 4 4.5 5 SVR p edic ed MOS Real MOS o SVR p edic ed MOS (Polynomial ke nel q=2) QP only PLCC = 0.8971 SROCC = 0.9218 RMSE = 0.4533 Fig. 10: Co ela ion be ween he second o de MOS alues and p edic ions p o ided by he SVM eg ession model o he QP only. 4. Conclusions and Fu u e Wo k In his con ibu ion, we ha e analyzed he possibili y o c ea e a da abase wi h a mo e ecen coding s an- da d (HEVC) om he exis ing anno a ed da abases o olde coding s anda d, i.e. JPEG, by using di e en objec i e measu es. While he esul s could no ye be alida ed by subjec i e es ing, some e idence was p esen ed ha he app oach is easonable. In addi ion, machine lea ning, in pa icula SVM, was used o e- la e he bi s eam in o ma ion o HEVC o subjec i e quali y. While his is a common app oach o aining No- Re e ence bi s eam models, ou wo k is ocused on un- de s anding he link be ween he encoding p ocess and he human pe cep ion. Fi s esul s conce ning he im- age pa i ioning used by he HEVC in a codec we e p esen ed in his pape . Fu he analysis o he encod- ing p ocess and he esul ing bi s eam in o ma ion is equi ed, also ex ending om s ill image o ideo in u u e esea ch. Acknowledgmen This publica ion is he esul o he p ojec imple- men a ion: Cen e o excellence o sys ems and se - ices o in elligen anspo II, ITMS 26220120050 suppo ed by he Resea ch & De elopmen Ope a ional P og amme unded by he ERDF. Re e ences [1] SEVCIK, L., M. VOZNAK and J. FRNDA. QoE p edic ion model o mul imedia se ices in IP ne wo k applying queuing policy. In: In- e na ional Symposium on Pe o mance E al- ua ion o Compu e and Telecommunica ion Sys ems (SPECTS 2014). Mon e ey: IEEE, 2014, pp. 593–598. ISBN 978-1-4799-5745-3. DOI: 10.1109/SPECTS.2014.6879998. [2] VAN WALLENDAEL, G., N. STAELENS, L. JANOWSKI, J. DE COCK, P. DEMEESTER and R. VAN DE WALLE. No- e e ence bi s eam- based impai men de ec ion o high e iciency ideo coding. In: 2012 Fou h In e na ional Wo k- shop on Quali y o Mul imedia Expe ience. Ya a Valley: IEEE, 2012, pp. 7–12. ISBN 978-1-4673- 0725-3. DOI: 10.1109/QoMEX.2012.6263845. [3] LEI, X., X. JIANG and X. MA. A no- e e ence ideo quali y e alua ion me hod based on HEVC bi s eam. In: Eigh h In e na ional Con e ence on Digi al Image P ocessing (ICDIP 2016). Chengdu: SPIE, 2016, pp. 1068–1074. ISBN 978-1-5106- 0503-9. DOI: 10.1117/12.2243822. [4] ANEGEKUH, L., L. SUN and E. IFEA- CHOR. Encoded bi s eam based ideo con en ype de ini ion o HEVC ideo quali y p e- dic ion. In: 2014 IEEE In e na ional Con e - ence on Communica ions (ICC). Sydney: IEEE, 2014, pp. 1296–1301. ISBN 978-1-4799-2003-7. DOI: 10.1109/ICC.2014.6883500. [5] ANEGEKUH, L., L. SUN and E. IFEACHOR. Encoding and ideo con en based HEVC ideo quali y p edic ion. Mul imedia Tools and Appli- ca ions. 2015, ol. 74, iss. 11, pp. 3715–3738. ISSN 1380-7501. DOI: 10.1007/s11042-013-1795-z. [6] SHAHID, M., J. PANASIUK, G. VAN WALLEN- DAEL, M. BARKOWSKY and B. LOVSTROM. P edic ing ull- e e ence ideo quali y measu es using HEVC bi s eam-based no- e e ence ea u es. In: 2015 Se en h In e na ional Wo kshop on Quali y o Mul imedia Expe- ience (QoMEX). Pylos-Nes o as: IEEE, 2015, pp. 1–2. ISBN 978-1-4799-8958-4. DOI: 10.1109/QoMEX.2015.7148118. [7] IZUMI, K., K. KAWAMURA, T. YOSHINO and S. NAITO. No e e ence ideo quali y assessmen based on pa ame ic analysis o HEVC bi s ea. In: 2014 Six h In e na ional Wo kshop on Quali y c 2019 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 443 INFORMATION AND COMMUNICATION TECHNOLOGIES AND SERVICES VOLUME: 17 |NUMBER: 4 |2019 |DECEMBER o Mul imedia Expe ience (QoMEX). Singapo e: IEEE, 2014, pp. 49–50. ISBN 978-1-4799-6536-6. DOI: 10.1109/QoMEX.2014.6982287. [8] HE, T., R. XIE, J. SU, X. TANG and L. SONG. A no e e ence bi s eam-based ideo quali y as- sessmen model o h.265/he c and h.264/a c. In: 2018 IEEE In e na ional Symposium on B oadband Mul imedia Sys ems and B oadcas ing (BMSB). Valencia: IEEE, 2018. ISBN 978-1-5386- 4729-5. DOI: 10.1109/BMSB.2018.8436879. [9] FAZLIANI, Y., E. ANDRADE and S. SHIRANI. Lea ning based hyb id no- e e ence ideo qual- i y assessmen o comp essed ideos. In: 2019 IEEE In e na ional Symposium on Ci cui s and Sys ems (ISCAS). Sappo o: IEEE, 2019, pp. 1– 5. ISBN 978-1-7281-0397-6. DOI: 10.1109/IS- CAS.2019.8702584. [10] ALIZADEH, M., A. MOHAMMADI and M. SHARIFKHANI. No- e e ence deep comp essed-based ideo quali y assessmen . In: 2018 8 h In e na ional Con e ence on Com- pu e and Knowledge Enginee ing (ICCKE). Mashhad: IEEE, 2018, pp. 130–134. ISBN 978-1- 5386-9569-2. DOI: 10.1109/ICCKE.2018.8566395. [11] REN, J., X. JIANG, L. XU and X. GUO. No- e e ence quali y assessmen o UHD ideos based on HEVC encoded bi s eam. In: 2017 4 h In e na ional Con e ence on Sys ems and In o ma ics (ICSAI). Hangzhou: IEEE, 2017, pp. 1319–1323. ISBN 978-1-5386-1107-4. DOI: 10.1109/ICSAI.2017.8248490. [12] HUANG, X., J. SOGAARD and S. FORCHHAM- MER. No- e e ence pixel based ideo quali y as- sessmen o HEVC decoded ideo. Jou nal o Vi- sual Communica ion and Image Rep esen a ion. 2017, ol. 43, iss. 1, pp. 173–184. ISSN 1047-3203. DOI: 10.1016/j.j ci .2017.01.002. [13] NAWADA, J., L. JANOWSKI and M. LESZCZUK. Modeling o quali y o ex- pe ience in no- e e ence mode. Jou nal o Telecommunica ions and In o ma ion Technology. 2017, ol. 2017, iss. 2, pp. 11–17. ISSN 1509-4553. DOI: 10.26636/j i .2017.114517. [14] NINASSI, A., P. LE CALLET and F. AUTRUSSEAU. Pseudo no e e ence image quali y me ic using pe cep ual da a hiding. In: Human ision and elec onic imag- ing XI. San Jose: SPIE, 2006, pp. 146–157. DOI: 10.1117/12.650780. [15] SHEIKH, H. R., Z. WANG, L. CORMACK and A. C. BOVIK. Li e image quali y assess- men da abase 2003. In: Labo a o y o Image &Video Enginee ing [online]. 2006. A ail- able a : h p://li e.ece.u exas.edu/ esea ch/Quali y/subjec i e.h m. [16] POMONARENKO, N., L. JIN, O. IEREMEIEV, V. LUKIN, K. EGIAZARIAN, J. ASTOLA, B. VOZEL, K. CHEHDI, M. CARLI, F. BAT- TISTI and C.-C. J. KUO. Image da abase TID2013: Peculia i ies, esul s and pe spec i es. Signal P ocessing: Image Communica ion. 2015, ol. 30, iss. 1, pp. 57–77. ISSN 0923-5965. DOI: 10.1016/j.image.2014.10.009. [17] ITU-R BT.500-13. Me hodology o he subjec i e assessmen o he quali y o ele ision pic u es. Gene a: ITU, 2012. [18] ITU-T P.910. Subjec i e ideo quali y assessmen me hods o mul imedia applica ions. Gene a: ITU, 2008. [19] HEVC. High E iciency Video Coding. Be lin: HHI, 2015. [20] SEVCIK, L., L. BEHAN, J. FRNDA, M. UHRINA, J. BIENIK and M. VOZNAK. P edic ion o subjec i e ideo quali y based on objec i e assessmen . In: 2018 26 h Telecom- munica ions Fo um (TELFOR). Belg ade: IEEE, 2018, pp. 1–4. ISBN 978-1-5386-7171-9. DOI: 10.1109/TELFOR.2018.8612127. [21] VQMT: Video Quali y Measu emen Tool. In: L’Ecole poly echnique ede ale de Lau- sanne (EPFL) [online]. 2019. A ailable a : h ps://mmspg.ep l.ch/ qm . [22] PINSON, M. H. and S. WOLF. An objec i e me hod o combining mul iple subjec i e da a se s. In: Visual Communica ions and Image P o- cessing 2003. Lugano: SPIE, 2003, pp. 583–593. ISBN 0-8194-5023-5. DOI: 10.1117/12.509909. Abou Au ho s Tomas MIZDOS was bo n in 1993 in Pop ad, Slo ak Republic. He ecei ed his B.Sc. deg ee in Mul imedia echnologies a he o Mul imedia and In o ma ion-Communica ion Technologies, Facul y o Elec ical Enginee ing and In o ma ion Technology, Uni e si y o Zilina in 2015. Cu en ly he is a Ph.D. s uden a he same depa men . His main a ea o in e es is Func ionali y and Quali y o Mul imedia Se ices, Quali y o Expe ience, Video Coding and Digi al Signal P ocessing. Ma cus BARKOWSKY ecei ed his D .-Ing. deg ee om he Uni e si y o E langen-Nu embe g c 2019 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 444