Cattle Body Intelligent Measurement based on Improved CenterNet.pdf

na

1. 3. Becca Systems 202 (2024-6472

a KB
ta

=

aa ty me ee nate e ml ot an cy,
(em ie eng ct eestor m per bs en pr ds
Cs a oped mb, un mer Dni o mer Ra 1011 er srt nd
oem eso move he tle eee yy rol De met pte o ve
‘ice scones enn ea pam cl a ay oe eT
‘Spm tnt epee! ip mee el ee ned oan opt mt tac
Se me nda nana mat ey en rey hema
{tes Cad aa re De ca th pl psy sd eg ye ete
‘ere tr rt ets amen ne

rad Care Baty Mere Cent, Kept Desc, nc Meer, Des ening

1 IemRonvenos

As th mia parameter e cate grow the Body size isan importas indicator o erlang the prove and
eal sts. Most o the toa ate body detection algorihuns ar man! measurement methods which
o oy require he sable cooperation o cts br also pet rede manual rn 0

The progress of acia! inteligence ecology in computer vision has promoted the development of
Heck body measurement, including cate bad measereat 0. For example, Zhu and Tono wed the
compte vision brary OpenCV to obtain be cade body ize dt, bi has poor ii and more ble
for spl secs 0. Zo mess be body sie tough e image daa ened y ie Kinet cama 0.
Zhou wed the muscle Retnex und Graus peros o cleat the key pois fte ap body ize 0
Although oboe a high acu, increased the compotion costa adios, the above mentioned
Berk have lirios uch a excessive computa resources ad law measurement eii, Wich ae
Ale o apply in practic.

‘The Deep Leaming Based on objet detection eros has als een apli in animal body measurement,
sich has uo categories Error! Reference source not fund. One isthe single-stage model represented by
SSD, YOLO, Cone aod Ctra where Centre complete de abet detection By imrodbeing cet
ky points based on ConerNet. Theaters two-stage mode represented by Faster R-CNN. Fr example, hno
Metal sed Mak RON or st boy detection which vas inproted by Faster RONNO, Aho the wc
stage models can achiev high accuracy y lso consume mare compaationa resources dut large munber
of arate. His worth menting th the estee single-stage js detection mel is tbe most advanced
single-stage model wich uses CNN o recognize ech objet asthe key pis and pays more anio e
infomation abot he central region of each object. Compared wth oer popular object detection modes,
Centre can model an objets the center ofthe Dounding box trough the ter polig operation, which is
sed fo egress on ther ais ofthe objet. This method an beter complete the detection of ey pois in
cate body image 0, Ualke general bjc detesta, cate body Keypoitdtstion in ae pose images i
comple sk, and he quality of etl pose image samples affects ate body keypoint detection or example,
lion aod cole changes ical busing images fete qualy of ale housing image Therefore. an
(ice kypoat crimiamton networks quie certe etc be key pois squid for at by

ever, he sizeof the central ra here te key points are located has a diet impact on the detection
res Ith ren st sal, the cl te of sal ets so nd thee isto arg, he ascracy ate
of are objets on, wich nt conde tthe alten ofthe nl key pin ofthe ctl body ie
region To obtain an accurate and sable set of key ois ile feature clusion technology mus be

‘Sait noma gene erh Urt Tesei, etn Cs
een ng
RE atone pomos

2. 3. Becca Systems 202 (2024-6472

wed replace radical fee cacaos mead performed by had. Due vo the arg discs ia
texte, color, and oct of key pints in cts, Bande heypoit modes is nt sable in tems of
‘nearest perlomance To sole the above proble, we we the basic menor ofthe Centre objet
tection mode replace the cviinal Reet 10 backbone network y DeseNet 10 base o Center, os.
to sor out the vanishing gradient sue cased by the many parameter o ReNel10. The network parameters
are reduced and the detection efiency is improved. At ls, te effectiveness of the improved ineligen
rca seo based on Ceres verd

TL Merioos

Numerous aces oe dentate Center apro perfomance in kypoin detection. and bees
sed in many pata ss. a ete body measurement ente uses het maps ad a single-level mode
are to detect and cli Key point of ae, which ave a hier recognition rt can lo reduce
computaonal os and improve measurement efcency. Based onthe above considerations we propose te
dy eliges measurement ee ar ies e DersNe fee exacto er Wil Center 10
dei the ey pois ol cale body in he ale stall sample. Compare with be aii Cte, ur
improved nero has bete detection perfomance fr he ey pois o an body size in can sal images,
hl inca ruca processing ine and improvi detecto eecy.

A Conos

Center as prove efect na vasc of computer soa tasks aná cn perfor bot object detection
and Kepoia locaton estimation tasks simultaneo. nation. Cer also excelent for eatle-body
Kespoi detection. The block disgram of Center is shown in Figur

Figue Block Dagam of Cesar

esta ses Resto ett ne fetes and detect object epi. Ree sles the probe of
incest dp ofthe network and degrada ie stckng ec fhe depe ewok 1 some! by
ing esl network othe etek Here we tke ResNe- 11 san example 1 analyze Ree

ResNet.101explily tres the sae infomation ofthe reins ye by ding a ink Moreover, Re
deity mapping tthe ret etw ad skips the operation of ose rmor yes Atte seins, aback
Propagation, he gradient of te next ayer o th network is dt puse 10 he presos lye, which solves
the problem ofthe disparo gai ofthe dep network. Honeve, he mue of parameters ofthe Reset
core sl age te taining tine i ag, an the hardware ree are rel hgh, The ResNet
cannot fly tie beige features, andthe ability tol he vaig sade! problem not god reg.
Figure 2 shows estr diagram of Ree 101 network

Figur 2 Secure Diagram of ResNet101 Neck
To solve te problems of ResNet10, we wea dense convolution network, nancy Dente o rplce
Reset Based on the original Centre, we replace ResNe- 101 with DensNe10UError! Reference source

3. 3. Becca Systems 202 (2024-6472

sot found. Compare with ResNe10, he etre extraction nemockDense-100 uses dense locks (Dese
Blocks. DB) tres the extract fetes ois bot necesa 1 flea ne edu ferro mappigs.
DenseNet101 wes the dimension coca operation to merge feo map. Ths ao only niche the fate
infomation and obins more feature maps, but also redes he convo options, wich significantly
reduces the neo parate, sven computations! cos, ad improve the detection eieny- Demo
ais mp dense blocks connected by convolutional ayers and pooling layers, DeaseNet has more
ple ass ht help compete to some exten forthe lack of seman formation ope
fee fon,
2. FoatneBuracten

enseNt 100 coins four ese modules nd bs be sme number o layers at ResNet-01, bu compared
to ResNet1O01,DeaseNet-00 hs fewer nero parameters ch compose of comoluional rel (CK) and
erage poling (AP 50 DeseNet 100 can save mare compton est. The network pater of DeseNet-
100 af led in Table

able 1: Newock parameters of DesseNe 10

ed
a | a me 2
CE ME ES :
ne cam a m
pe ‘8.8180 Fer 3
A | as ze =
= Le vu

‘DB ie base wi of Dense-10 as sbown In Figure Tie Tate map on ot
canes sf, ande future mp of level m1 ion as NVA, The nonlinear ansfomation A)
includes sever methods such a Bach Norman layer (BY), tentation fection REL, como
layer (Con) to minimiz e tot channels, anda 33 Con for keyoint eaangee The long oi ars
are use 1 represen dense connections, cometing 1 layers 0 as though asfomations A)
Filly te ouput of the +1 yer WN (1,920)

EI

igre Architect of Dense 100

A age number of densely connected module incrsses the numer of eae maps So, we use tation
layer o dia he dimension feature map, a fre exact, DenseNt-10 takes the sample with
tide ae of Rd, then pass it dom ad the detecte performs the next procsing sep.

© Detector
‘The heal image detector pers fat estimation on te features exacted rom Densent100 10.
callate central point of thecal body size espe group and the category information to wich he ale

nd size kpl group belongs. The ctrl keypinti be mio ofthe target bounding box ad i
called flows

o

4. 3. Becca Systems 202 (2024-6472

‘Where, ‘and ar the noise of the seta ey pints.) and fae the locations of he predicted
owasamping key pois. represents the aapive standard devin ofthe tre sz. € is the number of
categories and theres nly one category in his pape. à, represents the center ofthe anda keypain gon.
When parameter vale itis mad asthe te beis, itis marked athe backround

The size detecto isin charge of predicting the coniars ofthe fame. The fet detection bead peros
dovnsamplig on he pt samp, and he the cer pit fits bined by mining be ice
sor, Fly th coda oe enter poi ai fe rgrsin ae mapped a higher dimensional input
image

Centers an end en dsp leaming emo and we use amaia os method to improve poiioins
sccureyof key ares of at body size. Tots end, we we a mulas pealy L où ech sapling head
hic cale follows

Lente, o

Where, is th total los for Cone. a, % and À e he lores of them im, sie, and fet
pese À ad 7 ae constants 2701, 7-1, tn te tii proces the epoch is e as 200, andthe

D. 24 Cale Bod Data

Accurate determination ofthe size of each pr ofthe cts body is yt ect de cts growth
sans wich cn bea reference for beeling excelent eis oevaloting meat quality and Body weight The
schematic representation of cal body size dan can be found Figure.

Figue Schematic diagram of cane size dita

la Figure 4, we can ss ta teat and nal postions ofthe sight length of he cales body ar the
ional dance fom he ed ofthe soul the ear ee ofthe end ofthe chin the iil ad zl
patios o te height ofthe care’ body ae he vera distance om the highs pon of be ai tothe pound,
Ad e al nd ial postions ofthe length af the canes body ae the sight dance fo he end ofthe
shoulder tothe end o e tam,

OS

à Premi

Usually a sable dep lung nero cc ge amount o taining dt, but in most el word scene,
vis tical wo find enough high-quality wining dt ar ees be nds ofthe task. We refer 10 he inte
leasing proposed by Booaoag Error! Reference source not found 1 soi te above problem and improve
the detection accuracy, Taser leaning includes a source domai und à age dein. The specific proces is
eco a follows

CORTE NBC) o
‘Where. D) represents th source domain, D) represents tb pe domi. and A) represent he
eae sac od boundary probably distros spectively and Klar er

5. 3. Becca Systems 202 (2024-6472

Since our detection dt isthe camera image of eat poste, hich bas similar color and basi exe
«rats otal nage, 0 we have 1 pes the COCO dataset fs, use the pe said mode
eight a the na weights of our et, ile cur network, and finally Sine. the neck
rameter in tbe ining process. By wing transfer ang ita ot onl acelere the convergence speed of
‘Boel taining bat ao.

8. The Creation ofa Das

Them research object ofthis paper iste cate hasan image da collected fr. which contains
only one category of aes. The ae husbandy dataset coisa toa 1000 ages We vide the dataset,
iso a tii dust (600 images) anda tet das (400 image), and use data pretation techniques
iacadng random image rotation, mining ad random pruning augen the ninio ast, wich is
expanded to 3000 mages.

The cretion ofthe cane wining danse mail ives image acquisition and image processing Image
quis is caping images of cale hustuadey at diet distance and anges. Inge processing Is
increasing the amber of images cole. To ens elective taining ofthe model, the postion of the are of
Cane poste must be accureydetemiaed ase oo the inn image tems of etl postr Ia this werk, We
‘ed Labels t reste annotations forthe examples The nat generated alter being ae in an
XML fle, whieh coast ts. One he posi be Leypoir opi the standing ae ofeach cale,
td th lr ofthe keypi fom the center can be clued base the potion information. The teri
rectangular box rum onthe detected cate stad object In his work, we se the esp annotation 1 ete
minimum bounding box ofthe keypoit roo.
€. Baie decor

het images ee est bythe ined mode to sein he umber of cats, nd we daa boudin
Box around ach dette eae Ths, he mr of bounding boxes presents the number orecognize eats
inch test image. To check te effect of ou pop improved network, use precision. call e and FI
sear to evaluar the detection erfrmanc of can body size. The pein represents the acc ofthe
tried mode in pedia postive Samples, while the recall ate represas te accuracy of bw many ne
pie amples the ined mode dect The F sare vas developed o blanc hese o indicator Average
precision (AP) eval th shape ofthe precision and recl ate, defined as be average reson aa tof 1
fll paced real eves [001.02 . 1}. The precision cal rt, a FI score calcule follows.

wer
Ten
Pak

«

CET)

Wer, ste mabe of postive samples correct. FP isthe mbr of negative samples with
pasii samples incontables, she postive sample umber wih negative les coe abe.
2 iste acuncy te iste eat. Fn) isthe interpolation prison athe maximum precision. 7)
ist measurement precision a he ne of eal 7
D. Reuland anoto

ose dation performance e improved eto onthe cate poste ds, same images ons
hotest stare randomly selected as din objects. The detection esl are bow in Figure 5

6. 3. Becca Systems 202 (2024-6472

Figure Decio els of be network inthis paper.

From Figure $, ican be se hat the networkin this wok a fly dts he ey points required bythe
(are body size parameters. To bete sow te ect th keyoi group with he bounding bx of he red dei
line is site. However, thee ar sl some shoncomings, sich asthe posioia of some hey pits ofthe
seaplane pl, which canbe improve y collecting more image cta of ale bois inthe following ty

"o vale th measurement eet ofthe env: more compeebnsvly i this work, the experiment s
conduce with sel generated ts eof cate postures and compare With Faser. RCNN, original Cie,
tnd Centers .PWVSO. The compuso els ol detecting Le pins fr cnt body size re show in Table
2

“able 2 Compson Ress of Dire Neck

Same I pr ar ADM Fi] ECT
FaserRONN | 772 964 ES En 197
Cestemet | 787 965 wo 35.1 16
Tenemos PWV | 804 om A 2 me
Ou us os EN 94 us

By comparing wih ya cn a boy ment fn owe cn
be found a compared wih Faster RON, the rg Center and Centre PWYS, he of castrar
resul by our ae body measurement nero outperform he above tre advanced measurement networks in
Al even indicator, the detection effin is rely peeved andthe rang Ume red by 21
compared wth original Cente.

To fue ver the performance ofour neigen cat dy detection network we selected ates inthe
pasture and tally measure th cual body size prance. ben compared he cta measured vals wi
the measured wales call by ou proposed detection network. Theres ae shonin Table 3

Tale 3: Comparison Result of Body Sight Length

se Jem | Cenermeviem | Senos sen
No. [Measurement] Faser RCNNicm | ce as Our
TE miz 193 miz 10
2] 106 us 176 1396 1150
4] us me 1516 196 1336
5] iss ns 128 128 1106

7. 3. Becca Systems 202 (2024-6472

Table: Comparison Ress of Body Height

Nieiswene] Fate en | CER
x AAA Dunn
2 EEES ES ES 07
a; naz | 2 1352 mz maz
4 nso | 160 179 En 1947
s ms | ss 1255 120 en

“Tobie 3 Compan Re of ody Leah

No [Messremsat em] FatecRowNiem | Cemento | Cmenetpwvsen | ousen
7 nes EX ss us ss
2 102 192 1072 1567 1510
5 197 m7 127 197 En
+ 1503 103 EN EN ei
5 1566 146 126 1686 1560

‘The veil m Tabi 5 diate hat al ie mesures no ed in be experiment can ec
meas the body tag length. body bright. and body leg of cate, However, compared to aer
compare networks, the proposed neigen measurement bed on improved Cesta hs the highest
mesurer ocur, Sine te posiiaia of be sald canbe easy infenced when be et body length
Fs sit the measurement enor ls highest when the cath bad engi is sigh, but the detection eet best
te th ech Body ent os. The above experiment conti the eines of he proposed cate
nd elige measurement network, and we observe approximate acl mensreens in non-comac way.

To futher analyze he scuracy of dile easement ero, quatitaveeaknon was conduced
ing RMSE. MAE. and MAPE The calcio formulas ae as follows

ms [aus ©
mos ao ©
sure HE 0-20) o

mel 10 bal

ere, À represents te under of mensremens, represent the tl mumber of mesurent. 2,0)
represents the ah measurement vale. (ithe corresponding te vale.

Using formas (8) -(7) 1 acute a ol o 8 measurement eu in Tables 318. The calin ress
are showin Table,
“Toble6Evahton Ress

Indice | rasero | Centemet | Conner PWS oun
RMS EX] ns EJ 516
MABem EX] 1200 S10 276
MARES 1165 sor 36 199

From Table & we ca ee dar de proposed eigen mess sd on improved Cm las de
sales values inthe thc indicators of RMSE. MAE. and MAPE, the RMSE and MAPE ar oa 3.1621 and
1.99%, respectively, which fuer demons th the proposed ielignt cat de body measurement network

8. 3. Becca Systems 202 (2024-6472

can mare ass measure the sizeof cn, wile MAE is only 60m, inicaing tha he mesuenet
sure also higher stably

IV. Covewsion

‘Tosolse te cate body mesuremen problem a eliges ale body sz mensrrent neo based on
improved Cel proposed, which ss te mos advanced sgl tage obje tection model Center
ste bate network and replaces te ere extraction network RsNet witha more «ici network Dense
Moreover the collected sae o cate body image usd as an experimental ute valdte ur proposed
sort. The experimental esas sow nt ou cae body measurement network can ber adopt the ale
dy mage and achieve a good detection ect Compared withthe rial Centre, thecal body
measurement network accuse is significantly improved, ad iti significantly beter thn ober detection
mor, Our network provides technical spp for improin the precision level of animal usant and
promoting the development elige animal stad. Altboogh our experimental resul are excellent
there ar sil some shortcomings. For example, though tbe ctl body measurement network we propose as
high measurement accuracy, the repression of some key pons i tll nt accurate enough, inching te
posiioia o e key pits the Scapular ead I ure work, we conser collecting more images o ania
str poses dile seas salve more eliges messuremen problems o anil body Sie. A he
Lame te, we shoul lso improve te cae body measurement network, sli the eres acuacy of
Key pins and provide a measurement nor wi at convergence sped good subi, and high reson
for liz nia sde

AcKOWLEDENENT

he wt m tad la pr by Hin Pron Sess Tang Ras Pres
GE ON) Caryn ann an Fair Pre of Aseos Uy a
ein ot loa. Pv Sw Tc Rea Po CASINO md
I Resch Pre tg Eten Pn ns

ros

Dr rm i DE A 0
At

byw RES Dai ah ano gr ey amin EE
“nota cl to sl 0 PO Fa

prom at dni AR Cn

ta PE Da uc. 207 emi eed an Sing ip mr, E
u nt Se

LABS Za il al en ee Teto 4
ee

toy TR Ne u mc in” at Mat oe

(nin ly ti e Vb Hi a tl mii Cop
a man

(ash Ta a. Ks aie mt ig dr amor, 8
in Gs abet ong she

a elt ng nc

(mig uth fg Not mt rh ne ine
tent honda pte aye ges

tn in pa mn a u pen i npn vs
ano hs aap pe

(one a mE pf a ec pet i
e E sr ot hs pet AU

DRE trate names potas ty ei

(sone pi ty ct ni