Automated measurement of species and length of ﬁsh by computer vision

Fisheries Research 80 (2006) 203210

Automated measurement of speciesby computer visio

N.J.C, AberdN-5035

arch 2

Abstract

Trials of measare transpor g algomoment-inv 100%of 1.2 mm a f fish.research and ess up 2006 Else

Keywords: C

1. Introduction

Manual sorting of fish by species is carried out on all com-mercial anprocess is sciency onlabour on bcommerciaand weigh1986). Theing systemdata. This ronboard verequired fotraceability

Tayamabased on shfour specielar method

CorresponE-mail ad

species of fish. Utilising colour and shape parameters to sortfish by species, sorting reliabilities of 99% for 23 species offish have been achieved (Strachan, 1993a). However, these

0165-7836/$doi:10.1016/jd research fishing vessels (Strachan, 1994). Thelow and is the main limiting factor in terms of effi-commercial fishing boats and requires increasedoth types of ship. At present all fish caught by

l fishing vessels are manually graded by speciest in accordance with EC regulations (3703/85,re is a requirement for an automated fish sort-capable of recording species, length and weightequirement is driven by the need to reduce labourssels and to automate logging of the catch, which isr regulatory purposes and also to enable increased.

et al. (1982) describe a method for sorting speciesape and achieved a sorting reliability of 95% fors of fish. Wagner et al. (1987) describe a simi-and achieved a sorting accuracy of 90% for nine

ding author. Tel.: +47 45066042.dress: [email protected] (D.J. White).

results were obtained on sets of only 50 of each species offish and each fish took up to 15 s to process. In this method,for both roundfish and flatfish a grid is constructed whichdescribes their shape as a set of 36 elements. This works wellfor roundfish but not for fish with large aspect ratios suchas flatfish, for which it is possible that the width lines of thegrid cross over. Strachan (1993a) describes a simple shapegrid whereby the width lines are drawn parallel to each otherto avoid this problem. However, this method requires thatflatfish must be fed through the vision system with a certainorientation. This leads to increased complexity and cost inthe feeding systems.

Length measurements of fish are routinely taken onboardresearch vessels with an accuracy of 1 cm. This is doneeither using electronic measuring boards (Scantrol, Norway)or manual measuring boards where one person measures thefish and another notes the measurements, which are then laterentered into a computer. Both of these methods require eachindividual fish to be manually handled. Computer vision sys-tems that can automatically measure the length of fish in the

see front matter 2006 Elsevier B.V. All rights reserved..fishres.2006.04.009D.J. White a,, C. Svellingen b,a University of Aberdeen, Cruickshank Building

b Scantrol, Sandviksboder 1c, Bergen

Received 6 January 2006; received in revised form 30 M

a computer vision machine (The CatchMeter) for identifying andted along a conveyor underneath a digital camera. Image processinariant method, identify whether the fish is a flatfish or roundfish withnd species with up to 99.8% sorting reliability for seven species ocommercial ships is described. The machine can theoretically proc

vier B.V. All rights reserved.

omputer vision; Image processing; Fish; Fish species; Sortingand length of fishn. Strachan a

een AB24 3UU, UK, Norway

006; accepted 14 April 2006

uring different species of fish are described. The fishrithms: determine the orientation of the fish utilising aaccuracy, measure the length with a standard deviationThe potential application of the system onboard bothto 30,000 fish/h using a single conveyor based system.

204 D.J. White et al. / Fisheries Research 80 (2006) 203210

laboratory have been described (Arnarson and Pau, 1994;Strachan, 1993b) that give errors of less than 1 cm.

The aim of this work was to develop the next generation offish sorting equipment, employing modern hardware and pro-gramming techniques to identify species and measure lengthin real time. In particular this paper seeks to describe a methodwhich determines the orientation of fish based on the calcula-tion of the moments of a polygon, obtained from the silhouetteof fish. Fish can then be fed through the vision system in anyorientation on the conveyor belt, length measurements can bemade with errors of less than 1 cm and species is determinedbased on colour and shape. This will reduce the cost of anymechanical feeding systems and increase the throughput ofthe CatchMeter.

2. Materials and methods

2.1. Data

ImagesSea onboarAugust ofonboard thcalibrated bof the follopoglossoidSole (MicrGolden Re(Sebastes m

2.2. Mecha

The Catlight box (Fgens Lyngbtrolled bymain compalong the Vwere analy

Fig

2.3. Lighting and camera

Images of the fish, 1024 480 pixels with eight bits foreach of the red, green and blue channels were captured by aLumenera digital Universal Serial Bus (USB) video cam-era (Ottawa, Canada), with a VS Technology lens. Eachpixel represented a square of side 1 mm. To avoid specu-lar reflections from fish diffuse front lighting was suppliedby 150 W GE Tungsten Halogen bulbs. In order to facilitatethe fast identification of the silhouette of fish a semi trans-parent Volta conveyor belt was used with fluorescent bulbsunderneath. Tridonic dimmable high frequency ballasts (Jen-nersdorf, Austria) were used to ensure the fluorescent bulbsdid not flicker.

2.4. Computer and software

A desktop computer was used to run the C++ image pro-ng software with an AMD AthlonTM 64,3500+ Proces-GB of Corsair DDR400 RAM and an NVIDIA GeForcegraphics card. The compiler used was Microsoft Visual6.0. A flow chart describing how the software operateswn in Fig. 2.

g. 2. Flacquisition

of fish were obtained while testing in the Barentsd the Norwegian research vessel G.O. Sars during20051 and in the Eastern North Sea and Skagerate Dana during 20012. All images obtained werey the same method (Strachan et al., 1990). Imageswing fish were obtained: Long Rough Dab (Hip-es platessoides)1,2, Sole (Solea vulgaris)2, Lemonostomus kitt)2, Plaice (Pleuronectes platessa)1,dfish (Sebastes marinus)1, Deepwater Redfishentella)1 and Flounder (Platichthys esus)1,2.

nical handling system

chMeter (Scantrol, Bergen), including conveyor,ig. 1) and feeder were designed by Matcon (Kon-y, Denmark). The mechanical systems were con-

an Omron PLC (Kyoto, Japan) interfaced to theuter and software via an Ethernet link. Fish movedolta (Karmiel, Israel) conveyor belt at 1.5 m/s andsed by the computer vision system.

. 1. Schematic diagram of the CatchMeter system.

cessisor, 16600C++is sho

Fi ow chart showing basic steps in the fish analysis software.

D.J. White et al. / Fisheries Research 80 (2006) 203210 205

2.5. Colour calibration

Automatic colour calibrations were performed every hourto detect and correct for any variations in colour caused by thecamera or lights (Strachan et al., 1990). Briefly, this involvesa Macbeth colour chart (Baltimore, USA) being positionedunder the camera by an actuator controlled by the PLC andthen the colours being compared to reference measurements.Look-up tables were then automatically generated so thateach colour could be corrected for every image taken by thecamera.

2.6. Colour thresholding

Threshold colour values for the entire length of the con-veyor belt were also obtained over a time of 20 s every hour.The maximum and minimum colour values of the belt defineda region in the RGB colour space. Pixels outside this regioncould then be further analysed to determine the presence ofa fish in an

2.7. Findin

Singlelar to thescanned foels are fouare scanne

els withinthresholdther alongrequired nufish is assimage is csure the wis not anotto look forjoined togeimage.

Fig

2.8. Edge detection

The use of backlight to increase the contrast of the silhou-ette of fish

2.9. Rotati

After thestimationwhich twonot horizonwill be ina very acca sufficientrotation isthe fish in q

2.10. Orie

orders 15%ured (e head

Orie

r mos

arisonm offor fla

on be fish i

Roun

fter a fip andlated fsity ofe totalreate

fore athe fisod desan ex

due tand ion th

Flat

r thisf largeof a fiistant-sidedimage.

g a sh in an image

lines of pixels across each image perpendicu-direction of the conveyor belt are continuouslyr pixels outside the threshold values. If such pix-nd a small rectangular region around this pointd for more. Then if at least 40% of the pix-the region (10 10 pixels) are also outside thevalues then the process is repeated 2 cm fur-(Fig. 3). If both of these rectangles contain thember of pixels outside the threshold values, a

umed to be present and a 1024 480 bitmapaptured. This image is now scanned to makehole fish is contained within the image. If ither 1024 480 image is captured and scanned

the end of the fish. The two images are thenther so that the whole fish is contained within one

. 3. Determining the presence of a fish in an image.

Inwidthmeas

be th

2.11.

Focompbottocase

whiteof th

2.12.

Athe tocalcuintenof thwas gtherewisemethto bethemratelybased

2.13.

Fofish oedgeequida 100enabled Freeman (1974) chain code to be used.

on

e outline of the fish has been detected a roughof the rotation of the fish is obtained by findingpoints are furthest from each other. If the fish istal it is rotated until it is, so that any saved imagesa standard format. Although this does not giveurate measure of the orientation of the fish it is

input for the subsequent algorithms. The exactcalculated after the decision has been made thatuestion is a flatfish.

ntation of sh head and tail

to define which end of the fish is the head, twoof the length of the fish in from each end are

Fig. 5). The larger of the two widths was taken toof the fish.

ntation of sh back and belly

t roundfish the back is darker than the belly so abetween the average colour values of the top and

the fish identifies the orientation. This is not thetfish which tend to be coloured evenly on top andottom. This similarity in colour of the two halvess used to help identify flatfish.

dsh or atsh

sh has been split into two polygons representingbottom halves, the average greyscale intensity isor each area. If the difference between the averagethe top and bottom of the fish is greater than 15%possible intensity and the width to length ratio

r than 0.33, the fish was treated as being flat andcandidate for the new flatfish algorithms. Other-h was treated as a roundfish and processed by thecribed by Strachan (1993a). Redfish were foundception in that the flatfish code is better suited too their aspect ratio. These fish were treated sepa-dentified as candidates for the flatfish algorithmseir colour and aspect ratio.

sh grid

study a new method was developed for analysingaspect ratios (Fig. 4a) at any orientation. After thesh has been calculated (Fig. 4b) it is split up intolengths by 100 points which form the vertices ofpolygon (Fig. 4c). The moments (Appendix A)


Fig. 4. (a)

of this polynique (Straof area andrough indicused togethtail point otrue orientathe centralOriginal image of a plaice; (b) Outline; (c) 100 points on edge; (d) principal axis,

gon are then calculated using an integration tech-chan and Nesvadba, 1990). From this the centrethe principal axis are calculated. This axis gives aation of the orientation of the fish, which is thener with the chain code to find the optimal head andf the fish. This new axis is used to calculate thetion of the fish. Perpendicular lines are drawn to

line of the fish 15% in from each end, these widths

can then betail are (Fifish has beand three hgrid elemethe grid hasment (Fig.(10 equidiscentre of area and width lines; (e) grid; (f) fully processed.

used to decide which way around the head andg. 4d). After the rotation and orientation of theen decided 10 perpendicular widths are drawn inorizontal lines that split each segment into four

nts across the width of the fish (Fig. 4e). Whenbeen drawn the average colours in each grid ele-

4f) along with data regarding the shape of the fishtant width measurements along the length) can be


Fig. 5. Im

used to makset of equaby a canon

2.14. Colo

The avement are ca10 shapes)0 and 1. Shthe square rnormaliseddata were tutilising ca

2.15. Leng

The Cataccuratelydeformatiothe centraltip of the nmouth is gaat approximthe fish, thifour pointsmiddle axisperpendicuuntil enougend of the tin findingas both canroundfish.by the samcomputatioless in shap

2.16. Cano

Canonicing methodcanonical cables (specand with a cical discrim

the variables that maximizes the ratio of between-group andwithin-group variation. The canonical variates can then beevaluated and scores obtained so that unclassified observa-

can be assigned to a group whose mean score is closest.lassification score (Ci) for each species (i) can be cal-

ed for each fish from the linear combination of theification coefficients (Cij):jmaxj=1

CijQ(j) + Cio,

e i = 1,2 . . . n 1 (n is the number of species of fish), Cioonstant, Q(j) denotes the actual values of the variables

max is 124 (114 colour variables and 10 shape variables).btain average classification scores for each species ofthe system was first trained with 100 fish of each speciestested and then unclassified fish could be assigned to aes whose average was closest. Fish whose discriminantlay ouclassi

esults

ery fiidate f

top ar. Whsh fo

Whenincorrspecas un

ts.test t

Halibuotatioof 1.2a me

ith a m

1s for th

s

abage of a processed Dab showing length measurement line.

e an online classification by comparing them to ations that describe each trained species generatedical discriminant analysis.

ur and shape

rage red, green and blue colours in each grid ele-lculated and then all of the 124 (114 colours andvariables are normalized to give values betweenape variables are normalised by dividing them byoot of the total area of the fish, colour variables areby dividing them by 256. The colour and shape

hen processed to distinguish the species of fish bynonical discriminant analysis.

th

chMeter estimates length by obtaining a line thatdescribes the length of fish in any orientation orn. The line was drawn between eight points alongaxis of the fish as shown in Fig. 5. One point on theose or one between the upper and lower jaw if theping and one 10% along from this point. One pointately 60% and another at 70% from the head of

s is to avoid any belly flaps and fins. Then anotherin the tail of the fish, the first three are in theof the tail and the final point is found by scanning

larly away from the last point on the central axish data is acquired to choose an optimal point on theail. The tail and mouth require more computationthe optimal points to describe the overall lengthbe deformed in many orientations especially for

The length of both round and flatfish is obtainede method although in practice flatfish require lessn as they tend to be more symmetrical and deforme.

tionsThe cculatclass

Ci =

wheris a cand jTo ofish,to bespeciscore

were

3. R

Evcandratio,colou100 fified.were

seven

sifiedresul

Tolandand rationbeingtist w

TableResult

Specie

L.R. Dnical discriminant analysis

al discriminant analysis is a dimension reduc-derived from principal component analysis and

orrelation. Given a number of classification vari-ies), with a number of samples (fish) per speciesertain set of variables describing each fish, canon-inant analysis finds the linear combination of

SoleLemon solePlaiceFlounderD. RedfishG. Redfish

Total

Misclassifiedunknown andtside the 95% confidence interval for their speciesfied as unknown.

sh used in this test was correctly identified as aor the flatfish algorithms based on their aspectnd bottom half average greyscale intensity anden analysing the fish used to train the system allr each of the seven species were correctly classi-analysing the test sets only five deepwater redfishectly classified as golden redfish. For each of theies at least one and at most 12 fishes were clas-known, see Table 1 for a full breakdown of the

he length measurement algorithms a single Green-t was measured 100 times, in varying positions

ns. The length measurements had a standard devi-mm for the 413 mm fish (Fig. 6), the reference

ticulous manual measurement by a marine scien-easuring board.

e species sorting of flatfish in the Barents Sea

No. ofcalibration fishsorted correctly

Sortingreliability(%)

No. of testfish sortedcorrectly

Sortingreliability(%)

100/100 100.0 399/400(1) 99.8100/100 100.0 394/400(2) 98.5100/100 100.0 397/400(3) 99.3100/100 100.0 399/400(4) 99.8100/100 100.0 398/400(5) 99.5100/100 100.0 383/400(6) 95.8100/100 100.0 396/400(7) 99.0

100/100 100.0 2766/2800 98.8

and unknown fish: (15, 7) classified as unknown; (6) 12 asfive as Golden Redfish.


and Hal

4. Discuss

The resspecies ofwith largelimited thetem can bbeen showndard deviacurrent maEC regulat(EC, 1986)

The misof the fact tare not asIncreasingfish and incthis proble

For all sof each tesusually ariwas decideunknown aFurther opnumber ofrequired in

It was fdid not necimportant fing sizes aapproximaresent thisincreasingcation accusimilar in s

e decnalysiaper inalysis alsoase inshoulned bynd th

ghtingame instemsng theg acc

ded oin the

ing daFig. 6. Length measurement test for a single Greenl

ion

ults show that it has been possible to sort sevenfish with up to a reliability of 99.8%. Only fishaspect ratios were included in this study and thisnumber of species available, in practice the sys-

e trained with more species as required. It hasthat the system can measure length with a stan-

tion of 1.2 mm. This variation is far within theximum allowable error of 1 cm as dictated byions on research and commercial fishing vessels.classification of five deepwater redfish was a resulthat they were small (510 cm) juvenile fish, whichdistinctive in colour and shape as the adult fish.the size of the training set to include more juvenilereasing the resolution of the images could solve

Ththe athis pthe aIt wadecre

Itobtaisets athe lithe sthe sytreatisortinincluusedtrainm.even species of fish at least one and at most 12t set were classified as unknown. This situationses due to fish being deformed in some way. Itd that it was more acceptable to classify a fish asnd sort it out, rather than risk misclassification.timising the confidence levels could reduce thefish classified as unknown but more data would beorder to do this.

ound that increasing the size of the training setsessarily lead to better classification scores. Theactor was to include an even distribution of vary-nd shapes of each species of fish, 100 fishes wastely the minimum number of fish required to rep-range in the training set. Further tests showed thatthe number of species did not reduce the classifi-racy, but introducing different species that wereize and shape did.

order to enThe com

G.O. Sars,The systemmates previinformationtribution ofquickly andcan be usedget an accusystem is athe manuallimiting faas quicklyan entire camonitoring

The sof20100 msibut measured 100 times.

ision to use 114 colour and 10 shape variables ins was based on findings by Strachan (1993a). Int was shown that using only colour or shape data ins led to a reduction in the classification accuracy.shown that using fewer colour variables led to aclassification accuracy.

d be noted that the results shown in Table 1 wereincluding images from each source in the training

en classifying images from all sources. Althoughsetup and calibration method were designed to bethe 2001 and 2005 cruises, slight differences incould lead to a reduction in sorting accuracy whendata from each cruise as equivalent. Decreased

uracy could also occur when more species arer fish from only one sea area or time of year are

training set. There is a therefore a need to useta from different sea areas and times of year in

compass both seasonal and spatial variation.puter vision system is currently installed on the

where it will be used in routine fish stock surveys.is of value on research vessels since it auto-

ously manual processes and can present new catchthat is usually unavailable such as the colour dis-

fish and it can perform length measurements moreaccurately than current methods. This length dataalong with species and catch area information to

rate estimation of weight (Coull et al., 1989). Thelso of value on commercial fishing vessels sincesorting of fish by species is usually the efficiency-

ctor in terms of getting fish into storage freezersas possible. Since the machine can potentially logtch it may also be of use as a potential method fordiscards.

tware can fully analyse and classify each fish independing on the size using an AMD AthlonTM


64,3500+ processor. The limiting factor in the throughput ofthe machine is how fast the fish can be fed along the conveyor.Using a smooth Volta FHW2 belt it was found the maximumbelt speed tat a faster slarger >50the ship wathe system.for integratand will besystem wo3600 1 m fi

Before tfully implea number otem has beeso fish weretem must bwith a genebased on thbased on thpossibilityIt may be dis currentlyduced on sdesigned ansystem. Inveyor systebe circumvof edge detis to be incple fish inbe acquiredproblems mferent in siand sea areright or leftreat fish thspecies.

5. Conclu

A metha computedescribed.research veflatfish by sby the metwith a stanrecognise rby a similamachine isdesign pha10 cm fish/information

tial commercial value of the system is apparent but yet to berealised.

owled

e AuTotlaeir he

rrangirch vedvicer. TheNorwearch Ff Mar

ndix

omennitionobjec

; Bernomen

unninny twoistribued as

=

f(x, y)=x

e firsd M01and yhe cenM10M00

ey cainimum12

arct

rence

son, H.ape classificat

tein, H.usinghat is possible was 1.5 m/s. When the belt was runpeed the belt slipped under the fish, especially thecm fish. The maximum throughput achieved ons 7200 fish/h, which required two people feedingA mechanical feeding system has been designedion with the computer vision system by Matcontrialled in 2006. The introduction of this feeding

uld offer a theoretical maximum throughput ofsh/h or 30,00010 cm fish/h.he operation of the system is such that it could bemented on research and commercial fishing vesself issues still must be addressed. The feeding sys-n designed but was not tested during this sea trialfed onto the conveyor manually. The feeding sys-

e built and tested. Many flatfish are coloured on toprally white underside. These fish were classifiedeir coloured side. The feasibility of sorting themeir underside must be investigated along with theof a feeding system that could flip these fish over.esirable to reduce the size of the machine which3.5 m in length, especially if it is to be intro-

maller fishing vessels. A user interface must bed tested to allow non-expert people to operate theorder to facilitate integration with existing con-ms on ships the requirement for backlight mustented. Research into using more robust methodsection must be carried out for this. If throughputreased further the feasibility of analysing multi-one image must be investigated. More data must

and from different sea areas to investigate whatay arise given the fact that fish will appear dif-

ze and shape depending on age, sex, time of yeara. The feasibility of detecting whether flatfish aret sided should be investigated along with how toat have resulted through interbreeding of different

sions

od for automatic flatfish species detection byr vision pattern recognition system has beenThe machine is currently installed on a Norwegianssel the G.O. Sars. During testing it recognisedpecies with up to 99.8% accuracy for seven specieshod described in this paper and measured lengthdard deviation of 1.2 mm. The system can alsooundfish with an accuracy of approximately 99%r method (Strachan, 1994). The throughput of thelimited only by the feeding system which is in these and will offer a capacity of 3600 1 m or 30,000h. The machine is to be used to offer detailed catch

to scientists on the research ship and the poten-

Ackn

ThBjrnfor thfor aresea

his aMeteTheResetute o

Appe

Mrecogof an1984the mand C

Asity ddefin

Mpq

If

Mpq

Thsecon

the xfind t

xc =

Thof m

T =

Refe

Arnarshcla

Bernsbygements

thors would like to thank Jan-Tore vredal andnd of the Institute of Marine Research in Bergenlp and comments regarding this work and also

ng the test cruises on the Norwegian G. O. Sarsssel. Erik Andersen of Matcon in Denmark forregarding the mechanical systems of the Catch-companies and institutions that funded this work:gian Research Council, The Norwegian Fisheryund, The Nordic Council of Ministers, The Insti-ine Research in Bergen and Scantrol.

A. Moments

ts can be used in image processing and patternto calculate the principal axis and centre of areat described by a polygon (Bhanu and Faugeras,stein, 1986). Furthermore, it is possible to evaluatets of a polygon based on the vertices alone (Wilfgham, 1979; Strachan and Nesvadba, 1990).dimensional shape can be represented by a den-tion function f(x, y). The pq-th order moments are

xpyq dx dy.

is piecewise continuous then:y

xpyqf (x, y).

t moment M00 gives the area of the shape, theand third M03 give the moments of inertia about

axes, respectively. These moments can be used totre of area of the shape:

and yc = M10M00

.

n also be used to calculate the angle T of the axisinertia or principal axis of the shape:

an2(M00M11 M10M01)

(M00M20) (M00M02 M01) .

s

, Pau, L.F., 1994. Pdl-Hm morphological and syntacticssification algorithm real-time application to fish speciesion. Mach. Vision Appl. 7 (2), 5968.J., 1986. Determining the shape of a convex n-sided polygon2n + k tactile probes. Inform. Process. Lett. 22, 255260.


Bhanu, B., Faugeras, O.D., 1984. Shape matching of two-dimensionalobjects. IEEE Trans. Pattern Anal. Mach. Intelligence PAMI-6 (2),137156.

Coull, K.A., Jermyn, A.S., Newton, A.W., Henderson, G.I., hall,W.B., 1989. Length/weight relationships for 88 species of fishencountered in the North East Atlantic. Scot. Fish. Res. Rep.,43.

EC Commission Regulation 3703/85, 1986. Common Fisheries Policy,Community Grading Rules, a Guidance Note to the Fishing Industry.U.K. Fisheries Departments.

Freeman, H., 1974. Computer processing of line-drawing images. ACMComput. Surv. 6 (1), 5797.

Strachan, N.J.C., 1993a. Recognition of fish species by colour and shape.Image Vision Comput. 11 (1), 210.

Strachan, N.J.C., 1993b. Length measurement of fish by computer vision.Comput. Electron. Agric. 8, 93104.

Strachan, N.J.C., 1994. Sea trials of a computer vision-based fish speciessorting and size grading machine. Mechatronics 4 (8), 773783.

Strachan, N.J.C., Nesvadba, P., 1990. A method for working out themoments of a polygon using an integration technique. Pattern Recogn.Lett. 11, 351354.

Strachan, N.J.C., Nesvadba, P., Allen, A.R., 1990. Calibration of a videocamera digitizing system in the Cie L- Star U-Star Upsilon-Star colourspace. Pattern Recogn. Lett. 11 (11), 771777.

Tayama, M., Shimadate, N., Kubota, N., Nomure, Y., 1982. Applicationfor optical sensor to fish sorting. Reito (Tokyo). Refrigeration 57,11461150.

Wagner, H., Schmidt, U., Rudek, J.H., 1987. Distinction between speciesof sea fish. Lebesmittelindustrie 34, 2023.

Wilf, J.M., Cunningham, R.T., 1979. Computing Region Moments fromBoundary Representations. JPL Publication 79-49, Jet Propulsion Lab-oratory, California Institute of Technology, Pasadena, California.

Automated measurement of species and length of fish by computer visionIntroductionMaterials and methodsData acquisitionMechanical handling systemLighting and cameraComputer and softwareColour calibrationColour thresholdingFinding a fish in an imageEdge detectionRotationOrientation of fish head and tailOrientation of fish back and bellyRoundfish or flatfishFlatfish gridColour and shapeLengthCanonical discriminant analysis

ResultsDiscussionConclusionsAcknowledgementsMomentsReferences

Automated measurement of species and length of ﬁsh by computer vision

Documents

Transcript of Automated measurement of species and length of ﬁsh by computer vision