15
ins umen a ion iewpOin - 21 - MARTECH 23
ID05 MESOPELAGIC CRUSTACEAN HABITAT IDENTIFICATION AND ANALYSIS USING DEEP
LEARNING
An oni Bu gue a8, F ancisco Bonin-Fon 7, Damianos Cha zie angelou9
ABSTRACT
This pape p esen s a so wa e in as uc u e based on Deep
Lea ning aimed a iden i ying habi a s o ce ain species o c us-
acean ha colonize a eas o he ma ine Mesopelagic zone. De-
e mining hei p esense is done, in his case, om he de ec ion
o holes in he sand ha o m se s o bu ow s uc u es. P elim-
ina in e encing models a e ob ained om images cap u ed in
he No h Sea by awled UWTV (undewa e TV) s a ions, o e ing
qui e signi ican de ec ion success a ios.
Keywo ds – Deep Lea ning, Species Iden i ica ion, Image
P oocessing.
çThe g owing in e es in cap u ing species ha inhabi he Meso-
pelagic zone o he sea has o ced accu a e s udies and anlysis o
hei habi a s and habi s in o de o design sus ainable plans o
exploi a ion ha include adequa e ishing gea s [1] ha do no
des oy he en i e en i onmen and ben hos, and, in pa allel, e i-
cien p ese a ion ac ions o he species wi h highes comme cial
in e es . Neph ops no egicus is a c us acean e y alued in he
ish ma ke s, and a ade o be ween he desi e o s ake-hold-
e s o sa is y he con inuous demand and he p o ec ion o his
pa icula species is manda o y. Nowadays, adi ional s udies o
Neph ops no egicus habi a s and hy hms based on awling [2]
a e highly in asi e. Some al e na i es include he ideo eco ding,
day and nigh , om a g id o UWTV s a ions moun ed on sledges
owed by boa s, and ocus hei es ima ions in coun ing o bu -
ows, ei he manually [3] o wi h Deep Leaning [4]. Bu ows sa e
indi iduals om he awl ow cap u e, and, well iden i ied, a e
a clea sign o a colony. The ongoing spanish p ojec PLOME [5]
goes one s ep o wa d in he s udy o he Neph ops no egicus,
g abbing ideo sequences in si u wi h unde wa e ehicles, and
applying CNN o au oma ically de ec bu ows and animals o es-
ima e hei densi y.
Objec De ec ion (OD) is a compu e ision echnique aimed a
iden i ying and loca ing objec s in images. In he ecen yea s,
Deep Lea ning (DL) has shown ou s anding capabili ies o pe -
o m OD, clea ly su passing adidional me hods. One o he mos
p ominen DL-OD app oaches is You Only Look Once (YOLO). The
aim o his wo k is o ad ance in he au oma ic de ec ion o Ne-
ph ops no egicus bu ows in he con ex o he PLOME p ojec ,
expe imen ally e alua ing he abili y o YOLO 5 [6] o pe o m i
in unde wa e image y g abbed in ma ine en i onmen s densely
colonized wi h his species. This wo k ocuses on YOLO 5, hough
o he YOLO e sions could be es ed. T aining DL-OD sys ems e-
qui es la ge amoun s o labeled da a. In gene al his can be p ob-
lema ic because da a labeling is a edious and ime consuming
ask. When i comes o unde wa e image y, specially in deep sea,
he p oblem is magni ied since he da a i sel is sca ce and di icul
o ob ain. Da a augmen a ion alle ia es his p oblem bu i can
lead he DL-OD o o e i . Ou no el p oposal is o gene a e DL-
OD aining da a in he o m o image sub-samples ex ac ed a bi-
a ily om pho o-mosaics which we e buil om ac ual unde wa-
e images, ins ead o using he indi idual images o pe o m he
YOLO 5 aining. This app oach has se e al ad an ages. On he
one hand, i is less edious and e o p one o a human o ag a
single la ge image a he han hund eds o housands o smalle
images. On he o he hand, each objec in he mosaic is labeled
only once whe eas, i he indi idual images a e used, i has o be
labeled a e e y single image whe e i appea s. Ou app oach,
hus, a oids inconsis en labels among images and gene a es
mo e aining da a han he ac ual inpu images since iewpoin s
ha did no exis in he images can be ealis ically ex ac ed om
he mosaic. This app oach has also wo main d awbacks. Fi s , he
mosaic has o be cons uc ed and, second, some unde wa e a i-
ac s, such as igne ing o changes in illumina ion, a e emo ed by
he mosaic building ools. Ou p oposal o sol e his la e p ob-
lem is o a i ically add hese a i ac s when c ea ing he labeled
da a. The ully documen ed sou ce code o he p oposed da ase
gene a o is publicly a ailable a h ps://gi hub.com/abu gue a/
MOSAICDATASET.
EXPERIMENTAL RESULTS
A o al o 1810 bu ows we e hand labeled in i e di e en un-
de wa e mosaics o sizes anging om 1900x39603 pixels o
1900x50615 pixels. The pho o-mosaics we e supplied by he Func-
ioning and Vulne abili y o Ma ine Resou ces esea ch g oup o
he ICM (Ins i u de Ciences del Ma -Ba celona) and images used
o o m hem we e o iginally g abbed om a UWTV sys em o med
by sledges equipped wi h came as, and owed a a cons an speed
om a essel in he con luence o he Bal ic and he Scandina-
ian no h sea. All mosaics ep esen ec ilinea ansec s app ox-
ima ely 20 me e s long. Fou o hem ha e been used o build he
ain (4000 images) and alida ion (500 images) da ase s using ou
p oposal whils he i h mosaic was solely used o cons uc he
es da ase (500 images). All he gene a ed images ha e a esolu-
ion o 640x480 pixels and we e g ayscaled.
YOLO 5 MODEL INF. TIME [email p o ec ed] [email p o ec ed]:0.95
YOLO 5s 11.571 ms 0.797 0.422
YOLO 5m 28.181 ms 0.755 0.401
YOLO 5l 51.253 ms 0.792 0.427
Tab 1. Quali y me ics
16 ins umen a ion iewpOin - 21 - MARTECH 23
Fig 1. Image showing he in e ed bu ows.
A e wa ds, small, medium and la ge YOLO 5 a chi ec u es
(YOLO 5s, YOLO 5m and YOLO 5l) we e ained wi h he ain
da ase ine uning he hype -pa ame e s wi h he alida ion
da ase . Then, he quali y o each model was assessed on he
es da ase . The e alua ion was pe o med on a s anda d lap op
compu e (i7 CPU a 2.9 GHz) equipped wi h a NVIDIA GeFo ce
GTX 1650 and using o ch-1.11.0+cu113 o e Ubun u 20.03. Ta-
ble 1 summa izes he esul s. E en hough he la ge he model
he slowe he in e ence, he esul ing quali y is almos iden ical,
being close o a mean A e age P ecisions (mAP)@0.5 o 0.8 and a
[email p o ec ed]:0.95 o 0.4 in all cases. Figu e 1 shows a sample image
whe e in e ed bu ows appea in squa ed bounding boxes.
Since esul s sugges no majo di e ences in mAP be ween mod-
els, we decided o ocus on YOLO 5s because i shows a eason-
ably s able de ec ion a e o 86.423 ps. Figu e 2 shows, in he
le , he Recall-P ecision cu es o di e en le el combina ions
o In e sec ion o e Union (IoU) and A e age P ecisions (AP), and,
on he igh , he F1-Sco e changes depending on he IoU and he
con idence sco e h esholds used. This allowed us o de e mine
he op imal h esholds, which a e 0.1 o he IoU and 0.5 o he
con idence sco e. Using his op imal con igu a ion we eached no
only a p ecision close o 0.8 bu also a ecall o 0.74 and an F1-
Sco e o 0.77.
Fig 2. Recall-P ecision (le ) and F1-Sco e ( igh ) cu es o he ained YOLO 5s.
ACKNOWLEDGEMENTS
This wo k is pa ially suppo ed by G an PLEC2021-007525/AEI/10.13039/501100011033 unded by he Agencia Es a al de In es i-
gación, unde Nex Gene a ion EU/PRTR.
REFERENCES
[1] J. B čić, B. He mann, M. Mašano ić, S.K. Ši ne , and F. Škeljo, “CREELSELECT---A Me hod o De e mining he Op imal C eel Mesh: Case S udy on No way Lobs e
(Neph ops no egicus) Fishe y in he Medi e anean Sea”, Fishe ies Resea ch, ol. 204 , pp. 433-440, 2018.
[2] J. Aguzzi and F. Sa dÃ. A His o y o Recen Ad ancemen s on Neph ops no egicus Beha io al and Physiological Rhy hms. Re iews in Fish Biology and Fishe ies
18, 235–248 (2008).
[3] C. Lo dan, J. Doyle, R. Bunn, D. Fee and C. Allsop. A an, Galway Bay and Slyne Head Neph ops G ounds (FU17) 2011 UWTV Su ey Repo (2023).
[4] A. Nasee , E. N. Ba o, S. D. Khan and Y. V. Go dillo, “Au oma ic De ec ion o Neph ops no egicus Bu ows in Unde wa e Images Using Deep Lea ning,” Global
Con e ence on Wi eless and Op ical Technologies (GCWOT), Malaga, Spain, 2020, pp. 1-6.
[5] PLOME P ojec , h ps://plomep ojec .es/
[6] Ul aly ics, YOLO 5 h ps://py o ch.o g/hub/ul aly ics_yolo 5/, h ps://gi hub.com/ul aly ics/yolo 5