Machine Vision based Citrus Mass Estimation during Post-Harvesting Using Supervised Learning Algorithms

coatiarfΤεχνίτη Νοημοσύνη και Ρομποτική

17 Οκτ 2013 (πριν από 3 χρόνια και 10 μήνες)

85 εμφανίσεις

http://abe.ufl.edu/precag/
Machine Vision based Citrus Mass Estimation during
Post-Harvesting
Using Supervised Learning Algorithms
JunsuShin
a, Won Suk Lee
a, Reza Ehsani
b
aAgriculturaland Biological Engineering, University of Florida
bCitrusResearch and Education Center, University of Florida
Introduction
Amachinevisionsystemwasinvestigatedasameansofestimatingcitrus
fruitmassduringpost-harvesting.Anestimationoftheamountofcitrusfruit
canbemadebyanalyzingimagesthatcontainscitrusfruitattheprocessof
debrisremoval.Imageacquisitionforpreviousmachinevisionbasedyield
monitoringsystemtookplaceduringorbeforeharvesting.Thistypeof
sensingcouldleadtothedecreaseinestimationaccuracysincesensed
objectmightincludedebrisornon-fruitobjects.Hence,imageacquisition
aftertheremovalofthoseunnecessaryobjectswouldincreasethe
estimationaccuracy.Foraproposedmachinevisionsystem,imagesensing
afterthedisposalprocessofacitrusdebriscleaningmachinecorrespondsto
suchimageacquisition.
Objectives
•Developaimagesegmentationalgorithmbasedonsupervisedmachine
learningalgorithms.
•Designanddevelopamachinevisionsystemforcitrusmassestimationat
thetimeofpost-harvesting.
Materials and Methods
Hardware:
•CCDcolorcamerawithhighframerates(206fps)featuredcamera
•Incrementalencoderwasinstalledontherotatingaxisoftheconveyorfor
synchronization.
•Housingcoversthecameraandthelightningdevicesinordertoremove
theeffectofsunlightvariation.
•TwoExo-lights,DAQcard
Image acquisition
Image rectification
Fruit detection and
segmentation
using supervised
learning algorithm
Morphological
operations
& filtering
Total pixel area
calculation
Software:
Imageacquisition
•TwofieldexperimentsconductedatLykesgroveinFortBasinger,Florida
•Eachexperimenthasseveralsetswhichrepresentthedifferentyield
amount,abscissionusage,andharvestingconditions.
•Themachinevisionsystemcapturesimagesofcitrusfruitslidingoverthe
conveyorbeltofthecleaningmachine.
Set
number
Actual fruit mass
(kg)
Number of images
acquired
1st experiment11,492.32084
2979.8917
31,628.41947
2nd experiment12,004.92640
21,217.91787
31,510.52394
41,614.81733
Imagerectification
•Camera’sgeometryandlensdistortionmodelsarederivedthroughthe
processofcameracalibration.Twomodelsareusedtocorrectforintrinsic
deviationsandlensdistortions.
Fruitdetectionandsegmentation
•Binaryclassificationproblem:classifypixelsasfruitornon-fruit
•Supervisedmachinelearningalgorithmswereutilizedtosolvethebinary
classificationproblem.
•NaïveBayesclassifier:ParameterextractionusingMaximum
LikelihoodEstimation
•ArtificialNeuralNetwork:Multilayerfeed-forwardwith2hidden
layers
•DecisionTree:Pruningtoavoidover-fittingproblem
•Featurevector=[HueSaturationCbCr]
•Morphologicaloperationswithdisk-shapedstructuralelementwasusedto
removeanoiseandsegmentationerrors.
•Binaryimageasanoutputofthesegmentation
Results
Masscalibration
•Calibrationimagesetswhichconsistof8imagescontaining5fruit
sampleswithvaryingsizeandmassweretakenatthefield.
•Regressionanalysisonthepairsofmassandpixelareaforindividualfruit
•Estimated mass
kg

pixel area

Imagesegmentation
Mappingequation
Experiment
number
Error sum
of squares
(SSE, kg)
Coefficient of
determination
()
RMSE
(kg)


10.03220.9240.02910.0000686-0.0273
20.03030.9290.02820.0000718-0.0659
Experiment
Set
number
Actual
mass
(kg)
Naïve BayesNeuralNetwork
Decision Tree
Estimated mass (kg)
Error
(%)
Estimated mass (kg)
Error
(%)
Estimated mass (kg)
Error
(%)
1st
11,492.31,549.4-3.81,560.7-4.51,346.29.8
2979.8855.212.7883.79.8979.30.1
31,628.41,599.71.81,605.81.41,414.713.1
2nd
12,004.92,222.8-10.82,197.5-9.62,222.2-10.8
21,217.91,224.5-0.51,250.2-2.61,324.0-8.7
31,510.51,590.9-5.31,597.8-5.71,694.2-12.1
41,614.81,427.311.61,430.011.41,479.28.4
Estimated mass
kg

pixel area

RMSE (kg)R-square
Naïve Bayes121.70.929
Artificial Neural Network120.50.924
Decision Tree187.10.804
Massestimation
•Regressionanalysisbetweentheactualfruitmassandthefruit
estimatedmass
Problems
•Fruitmissingoroverlappinginimagecapturing
•Synchronizationproblem:varyingconveyorspeed,sensornoise
•Fruitoverfeeding,housingblockingthecameraviewinpart
Conclusions
•Amachinevisionsystemforcitrusmassestimationduringpost-
harvestingwasdesignedandimplemented.
•Fortheimagesegmentation,pixelclassificationalgorithmswere
implementedbasedonsupervisedmachinelearningalgorithms,such
asnaïveBayes,artificialneuralnetworkanddecisiontree.
•ThenaïveBayesandtheneuralnetworkmodelperformedbetterthan
thedecisiontreemodelinpixelclassification,whichledtoR-square
morethan0.92inconductingmassestimation.Evenifconsidering
ratherhighRMSE,itcanbeconcludedthatthemassestimationwas
performedreasonablywellusingthetwomethods.
Fig. 1. Schematic image of citrus
debris cleaning machine
Fig. 2. Housing
Fig. 3. CCD cameraFig. 4. Incremental encoder
Fig. 5. Block diagram of image processing algorithm
Table 1. Field experiment summary
Fig. 6. Image taken during
the field experiment
Fig. 7. Binary image created using
naïve Bayes classifier
Fig. 8. Binary image created using
Artificial Neural Network
Fig. 9. Binary image created using
Decision Tree
Table 2. Results of regression analysis on the calibration sets
Table 3. Summary of mass estimation results
Table 4. Comparison of mass estimation results