资源目录
压缩包内文档预览：
编号：21461358
类型：共享资源
大小：841.10KB
格式：ZIP
上传时间：2019-08-13
上传人：机****料
认证信息
个人认证
高**（实名认证）
河南
IP属地：河南
12
积分
- 关 键 词：
-
外文
翻译
评估
不同
模式识别
苹果
技术
分类
- 资源描述：
-
外文翻译-评估不同的模式识别苹果的技术分类,外文,翻译,评估,不同,模式识别,苹果,技术,分类
- 内容简介：
-
Available at journal homepage: /locate/issn/15375110Research Paper: PHPostharvest TechnologyEvaluation of different pattern recognitiontechniques for apple sortingI. Kavdra,?, D.E. GuyerbaDepartment of Agricultural Machinery, C - anakkale Onsekiz Mart University, 17020 C - anakkale, TurkeybDepartment of Biosystems and Agricultural Engineering, Michigan State University, East Lansing, MI, USAa r t i c l e i n f oArticle history:Received 4 November 2006Accepted 24 September 2007Available online 28 November 2007Golden Delicious apples were classified using parametric and non-parametric classifiersinto three quality classes. The features used in classification of apples were hue angle (forcolour), shape defect, circumference, firmness, weight, blush percentage (red natural spotson the surface of the apple), russet (natural netlike formation on the surface of an apple),bruise content and number of natural defects. Different feature sets including four, five andnine features were also tested to find out the best classifier and feature set combination foran optimal classification success. The effects of using different feature sets and classifierson classification performance were investigated. The feature set including five featuresproduced slightly better classification results in general compared to feature sets includingfour and nine features. When the classifiers were compared, it was determined that themulti-layer perceptron neural network produced the highest classification results (up to90%) while 1-nearest-neighbour and 2-nearest-neighbour classifiers followed this classifierwith an 81.11% classification success. The 3-nearest-neighbour and decision tree classifiersresulted in similar classification success (75.56%). The parametric plug-in decision ruleclassification resulted in the lowest classification success. Principal component analysisand linear discriminant analysis techniques were applied on the training data with nine,five and four features to visualise the degree of separation of the three quality classes ofapples. As a result of this application, some improvements were observed in separation ofthe three quality classes from using four input features to nine features especially usingprincipal components although some overlaps still existed among the classes.& 2007 IAgrE. Published by Elsevier Ltd. All rights reserved.1.IntroductionComputer-generated artificial classifiers that are intended tomimic human decision making for product quality haverecently been studied intensively. Combined with high-tech-nology handling systems, consistency is the most importantadvantage the artificial classifiers provide in classification ofagriculturalcommodities.Inaddition, the advantagesofautomated classification operations over conventional manualsorting operations are objectivity, null or low labour require-ments and reduction in tedious manual sorting.Patternrecognitiontechniqueshavethecapability ofimaging the distribution of quality classes in feature space.As a result, for some time different pattern recognitionalgorithms have been studied for classification of agriculturalproducts. The number of features plays a key role indetermining the efficiency of the pattern classification interms of time and accuracy.ARTICLE IN PRESS1537-5110/$-see front matter & 2007 IAgrE. Published by Elsevier Ltd. All rights reserved.doi:10.1016/j.biosystemseng.2007.09.019?Corresponding author.E-mail addresses: kavdiris.tr (I. Kavdr), guyer (D.E. Guyer).B I O S YS T E M S E N G I N E E R I N G99 (2008) 211 219The aim of this study was to classify Golden Deliciousapples using six different parametric and non-parametricpattern recognition classifiers and different quality fea-turesetsfordeterminingtheoptimalclassifierandfeature set combination. Apples were first graded by a humanexpert who followed USDA standards for apple qualityandexpectations.Followingthis,thesamesetsofapples were classified by the artificial classifiers that weredeveloped.2.Literature reviewAlchanatis et al. (1993) developed and applied three differentclassification algorithms in sorting tissue culture segmentsusing machine vision. The features used in classificationswere colour values belonging to different plant parts. The fivecolour values were the averages of red, green and blue colourcomponents, the area of the dark green region and the area ofthe light green region. Images of plant segments wereclassified into two groups as good or bad by each of the threeclassifiers. It was stated by Alchanatis et al. (1993) that a goodclassifier should be both fast and accurate. Three classifiersinvestigated weretestedforthreeproperties.Thefirstclassifier, the maximum log-likelihood, was a parametricclassifier which used the colour parameters as features. Thisclassifier assumed that features were normally distributedand independent of each other. Probabilities of a good plantletand a bad plantlet were estimated using the colour feature,andtheirmeans andstandarddeviations.Thesecondclassifier used was a non-parametric K-nearest neighbour(K-NN) which assigns the pattern the label that is mostfrequently represented among the K-nearest samples in thetraining set (Alchanatis et al., 1993). The third classifier was aneural network (NN) (Boltzman recognition network) classi-fier. NNs are known for their potential for high accuracy and,with their parallel computation structure, rapid execution.However, it may take a long time to train NNs. It wasconcluded by Alchanatis et al. (1993) that NN and machinevision provide a practical system for the separation of goodpotato plantlets from bad ones. Colour provides a significantadvantage for image recognition.Miller et al. (1998) evaluated different pattern recognitionmodelsforspectralreflectanceassessmentofappleblemishes.Multi-layerbackpropagationNN,unimodelGaussian, K-NN and nearest cluster algorithms were devel-oped and tested. Multi-layer back propagation NN pro-vided the highest classification (up to 96% and 85% in thefirst and second years, respectively) results with minorvariations for different number of hidden nodes and NNtopology.Chen et al. (1997) obtained improved classification results bypreparing input data to NNs using principal componentanalysis (PCA). Wholesome poultry and unwholesome poultrycarcasseswereclassifiedintotwooutputgroups.PCAprovided a powerful input feature set to the NN classifier.Input features extracted through PCA analysis resulted inhigher classification success in general, except in one case,compared to using original features.Classification procedures were developed using NNs andstatistical classifiers for sorting potatoes and cereal grains(Kirsten et al., 1997; Luo et al., 1999); in both studies, non-parametric classification approaches performed better com-pared to statistical methods although the difference was notsignificant in the potato classification study (Kirsten et al.,1997).Kim et al. (2000) applied linear and non-linear recognitionmodels for classification of fruit. Various feature extractionand dimensionality reduction techniques were performed onthe spectral data obtained from visible and near-infraredspectra. Linear pattern recognition techniques such as lineardiscriminant analysis (LDA) and non-linear techniques basedon multi-layer perceptrons (MLPs) were used to classify theproducts. In the results, non-linear approaches producedsuperior classification results.Penzaetal.(2001)usedpatternrecognitiontechni-questoclassifyfood,beveragesandperfumes.Suc-cessful results were obtained with PCA and cluster analysismethods.Leemans et al. (2002) developed an on-line fruit gradingsystembasedonexternalqualityfeaturesofapplesusingquadraticdiscriminantanalysisandNNs.Bothgrading algorithms resulted in similar results (79% and 72%)for both varieties studied. Similarly, Hahn et al. (2004)used discriminant analysis and NNs to detect Rhizopusstolonifer spores on tomatoes using spectral reflectance.The NN classifier outperformed the discriminant analysisapproach.ARTICLE IN PRESSNomenclaturednumber of featuresdEEuclidean distance between patterns i and kenumber of misclassified patternsgm, gk,discriminant functions for classes m, and ki, kpatterns, the distance between which ismeasuredMnumber of classesmjmean of the jth featureNtotal number of samplesP(wm)priori probability for class mSjstandard deviation of the jth featurewmclass mxijoriginal value of the jth feature of ith patternxijnormalised value of the jth feature of ith patterneestimate of true error ratemimean of class iSicovariance matrix for class iSubscriptsiindex for the test patternsjnumber of featureskindex for the training patternsB I O S YS T E M S E N G I N E E R I N G99 (2008) 211 2192123.Materials and methods3.1.Data acquisitionNine features were measured from Golden Delicious apples.These were hue angle (for colour), shape defect, circumfer-ence, firmness, weight, blush (red natural spots on the surfaceof the apple) percentage, russet (a natural netlike formationon the surface of an apple), bruise content and number ofnatural defects. Firmness was measured using a Magnes-sTaylor (MT) tester applying an 11mm diameter probe intoabout an 8mm depth (Effegi-McCormick, Yakima-FT-327).Colour was measured using a CR-200 Minolta colorimeter inthe domain of L, a and b, where L is the lightness factor and aand b are the chromaticity coordinates (Ozer et al., 1995). TheHue angle (tan?1(b/a), which was used to represent the colourof apples, was shown to be the best representation of humanrecognition of colour (Hung et al., 1993). The sizes of thesurface defects (natural and bruises) on apples were deter-mined using a special figure template, which consisted of anumber of holes of different diameters. In addition, a shapedefect (lopsidedness) was measured using a Mitutoya electro-nic calliper (Mitutoya Corporation) and taking the ratio of themaximum height of the apple to the minimum height. Themaximum circumference was measured using a Crantoncircumference measuring device (Cranton Machinery Co.).Weight was measured using an electronic scale (Model no:CT1200-S serial no: 3403, capacity 120070.1g). Programmingfor the classifiers was done in Matlab.3.2.Data collection and handlingThe number of apples used for each class was determinedbased on the availability of specially featured apples in the setof apples collected for this study. The total number of appleswas 181 which included three classes as bad (class-3),medium (class-2) and good (class-1) quality. The size of thepattern matrix was 181?9 where nine represented thenumber of features. Eighty of the apples were kept at roomtemperature for four days after harvest while another 80 werekept in a cooler (at about 31C) for the same period, withoutapplying any quality pre-sorting, to create colour variation onthe surfaces of apples. In addition, 21 of the apples wereharvested before the others and kept for a further 15 days atroom temperature for the purpose of creating a variation inthe appearance of the apples to be tested. Apples were gradedfirst by a human expert and then by the classificationalgorithms developed. The expert was trained on the externalquality criteria of apples for good, medium and bad applegroups defined by USDA standards (USDA, 1976). The USDAstandards for apple quality explicitly define the qualitycriteria so that it is quite straightforward for an expert tofollow up and apply them. Extremely large or small appleswerealreadyexcludedbythehandlingpersonnel.Apples were graded by the human expert into three qualitygroups depending on the experts experience, expectationsand USDA standards (USDA, 1976). The numbers of applesdetermined for each quality group by the human expert aregiven in Table 1. The low number of apples for class threeprovided a disadvantageous situation.The same set of apples was classified using fuzzy logic in aprevious study (Kavdir & Guyer, 2003).3.3.Classification algorithmsDifferent classification algorithms were applied to classifyapples after measuring the quality features and grading of theapples by the expert. The performance of the classifiers wascompared with one another and the expert.Parametric and non-parametric classifiers were used. Para-meter estimation in the parametric approach was done usingmaximum likelihood estimation. Estimated parameters werethen put in the plug-in decision rule (PDR). For the non-parametric approach, the K-NN decision rule with differentvalues of K (1, 2, 3), a decision tree (DT) classifier which wasformed in the S-plus (Venables & Ripley, 1994) statisticalpackage program using tree function and a MLP NN classifierusing NNs Toolbox in Matlab were used. Classifiers weredeveloped and tested on a Pentium Centrino M 760 2.0GHzpersonal computer with 1GB of RAM.3.3.1.Pre-processing of data and feature selectionEach feature was normalised using Eq. (1) below to eliminatethe unit difference between themx0ijxij? mjSj,(1)where xijis the original value of the jth feature of the ithpattern, x0ijthe normalised value of the jth feature of the ithpattern, mjthe mean of the jth feature, Sjthe standarddeviation of the jth featuremj1nXni1xij(2)andSj1nXni1xij? mj2.(3)Three feature sets were used in the classification applica-tion. Two different subgroups (with four and five features)oftheoriginalninefeaturesandtheoriginalninefeatures were used as input feature sets to the classificationalgorithms. The aim was to find an optimum feature setARTICLE IN PRESSTable 1 Number of samples used for training andtestingClassNumber oftraining dataNumber oftesting dataTotalClass-16464128Class-2222244Class-3549Total9190181B I O S YS T E M S E N G I N E E R I N G99 (2008) 211 219213that results in an improved classification and a shorterclassification period.PCA and LDA techniques were applied on the data withfour, five and nine features to visualise the degree ofseparation of the three quality classes of apples (Figs. 13).The training data devoted for each class were used in S-Plusfor this application.No specific feature selection algorithm was applied in thestudy. The first classifier to be developed was the decision treeclassifier formed in S-Plus (Venables & Ripley, 1994). Thetraining data for three quality groups with originally mea-sured nine features were submitted to the decision treeclassifier. In this procedure, the decision tree classifier withthe built-in feature it had chose its own feature sub set thatwas effective in classification. As a result, the decision treeclassifier determined four features, which were colour, shapedefect, weight and russeting, as effective features eliminatingthe remaining five features. Later on, firmness, which is animportant feature in fruit quality, was added to the fourfeatures determined by the decision tree classifier. Therefore,three feature sets each having four, five and nine features(including all those measured) were formed and used inclassifications performed by six classifiers.The following parametric and non-parametric classifierswere used in classifications with three different feature sets.3.3.2.Plug-in decision ruleParameter estimation in this parametric approach was doneusing maximum likelihood estimation. Estimated parameterswere then put in the plug-in decision rule. The data wereassumed to have a multivariate Gaussian distribution, whichis given as follows:pxi1;xi2;.;xid12Pd=2jSij1=2exp ?12x ? miTS?1ix ? mi?,(4)where d is the number of measurements and the expres-sion expf?12x ? miTS?1ix ? mi?g is the Mahalanobis distanceARTICLE IN PRESSsecond discriminant variable4202101GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGMMMMMMMMMMMMMMMMMMMMMMBBBBBsecond principal component2042102GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGMMMMMMMMMMMMMMMMMMMMMMBBBBB232first discriminant variable2first principal component1Fig. 1 Visualisation of the 3 quality classes of apples in two-dimensional projections of a four-dimensional feature space;the training data were transformed using both linear discriminant analysis and principal component analysis, and then thefirst 2 linear discriminant variables (and first 2 principal components) were plotted for quality groups (G, good (class-1)quality apple; M, medium (class-2) quality apple; B, bad (class-3) quality apple).second discriminant variable420first discriminant variable2102GG GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGMMMMMMMMMMMMMMMMMMMMMMBBBBBsecond principal component20432102GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGMMMMMMMMMMMMMMMMMMMMMMBBBBB2first principal component2131Fig. 2 Visualisation of the 3 quality classes of apples in two-dimensional projections of a five-dimensional feature space; thetraining data were transformed using both linear discriminant analysis and principal component analysis, and then the first2 linear discriminant variables (and first 2 principal components) were plotted for quality groups (G, good (class-1) qualityapple; M, medium (class-2) quality apple; B, bad (class-3) quality apple).B I O S YS T E M S E N G I N E E R I N G99 (2008) 211 219214measurement that, together with the exponential componentin the Gaussian distribution function, makes the probabilityof a pattern small if it is far from the mean.Priori probabilities of p(w1), p(w2) and p(w3) were determinedusing training samples from each class. Unknown parametersof miand Si, which are the mean and covariance matrix forclass i, were estimated from training samples using thetechniqueofmaximumlikelihoodestimationgivenasfollows:mi ?1nXnixi,(5)Si ?1niXnix ? mi2.(6)Using these parameters, the discriminant function for eachclass was calculated using Eq. (7):gixi ?12xi? miTS?1ixi? mi ?12ln Si? ln Pwi.(7)Using the value of gi(xi) in the decision rule given below, thetest pattern was assigned to the class whose discriminationfunction output the highest value. The discrimination func-tion gm(xi) for class wmis given as follows:Assign xito class wmif gmxi4gkxifor all mak,(8)where m and k are the classes for xito be assigned.3.3.3.K-nearest-neighbour (K-NN) classifierThe nearest-neighbour classifier is a non-parametric classi-fier which does not make any assumptions on the form of theconditional density of a class and assigns the pattern the labelmost frequently represented among the K-nearest samples inthe training set (Alchanatis et al., 1993). K represents thenumber of nearest neighbours. To assign a test pattern aclass, the entire training data are used, i.e. distances betweenthe test pattern and each of the training patterns have to bemeasured. Although training of this classifier is quite simple,it requires a large computer memory as it has to keep theinformation of every sample in the training set. Variousmetricsareusedinmeasuringthesimilaritybetweenpatterns. The Euclidean distance was used to measure thedistance (similarity) between patterns in this study. In the1-NN classifier, the test pattern was assigned to the class,which contained the pattern (training) that was closest to themeasured pattern (testing). Also, 2-NN and 3-NN classifierswere tested; in these classifiers, the pattern was assigned tothe class which had the majority of patterns represented inK-NN. When each of the nearest-neighbour patterns (training)was from a different class in 2-NN and 3-NN classifiers, thetest pattern was assigned to the class that had the closestmember to the test pattern, i.e. 1-NN classification procedurewas applied when each nearest-neighbour pattern was from adifferent quality class in 2-NN and 3-NN classifiers. Theeuclidean Distance is expressed as follows;dEi;k Xdj1xi? xk224351=2,(9)where d is the number of features, dEis the Euclidean distanceindex between the patterns i and k, xiis the position of thetest pattern and xkis the position of the training pattern.3.3.4.Decision tree (DT) classifierThenon-parametricandhierarchicalDTclassifierwasformed by splitting of subsets of the training data intodescendant subsets. Binary tree structure was used to formthe DT which was generated in S-plus (Venables & Ripley,1994). In this technique, the final variables (effective features)and the thresholds for the leaf nodes (for selected features)were automatically determined in designing the DT using allavailable features (training set) as input to the system. The DTclassifier had a feature selection algorithm built in S-plus.Implementing the DT classifier generated in S-plus (Venables& Ripley, 1994) produced an appropriate subset of fourfeatures (colour, shape defect, weight, russeting) to separatethe data into three quality classes. Selected features and theirthresholds, determined for training data by using the treefunction in S-plus, were later used in implementing the DTARTICLE IN PRESSsecond discriminant variable864202GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG GGGGGGMMMMMMMMMMMMMMMMMMMMMMBBBBB2second principal component202321GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGMMMMMMMMMMMMMMMMMMMMMMBBBBBfirst discriminant variable024012first principal component46Fig. 3 Visualisation of the 3 quality classes of apples in two-dimensional projections of a nine-dimensional feature space;the training data were transformed using both linear discriminant analysis and principal component analysis, and then thefirst 2 linear discriminant variables (and first 2 principal components) were plotted for quality groups (G, good (class-1)quality apple; M, medium (class-2) quality apple; B, bad (class-3) quality apple).B I O S YS T E M S E N G I N E E R I N G99 (2008) 211 219215classifier in Matlab for classification of the testing data. Thefeatures selected in the formation of the DT classifier werealso used in the rest of the classifiers. Detailed information ontree partitioning methodology can be found in Chapter 13 ofVenables and Ripley (1994).The DT classifier was used only with the feature setincluding four features. The rest of the classifiers were usedwith all the feature sets including four, five and nine features,respectively.3.3.5.Feed-forward neural networks classifier (multi-layerperceptron)The most important feature of NNs is their learning abilityfrom inputoutput relationships. However, most populartypes of NNs need to be retrained when new patterns(features) are added to the training set. Long training periodsand unclear procedure of how NNs work are their maindrawbacks.The MLP is the most widely used type of NNs. MLP has asimple structure; it consists of an input layer with a numberof neurons equal to the number of features, hidden layer(s)and an output layer. With no hidden layer, the perceptron canonly perform linear tasks.In this study a four-layered perceptron (including the inputlayer) with two hidden layers was used; the number ofneurons in the input layer were four, five and nine (all thefeatures measured), respectively, for three different sets offeatures (Fig. 4); and the number of neurons in the first andsecond hidden layers was eight and four, respectively: thenumber of hidden layers and the number of neurons in thehiddenlayerswereselectedbasedontrialanderror.Experiments started with using one hidden layer first, usingdifferent number of neurons in it. Later, trial and errorexperiments continued using two hidden layers in thestructure of the NN with different number of neuronsin it. Eight neurons in the first hidden layer and four neuronsin the second hidden layer produced the optimal resultscompared to the rest of the trial and error experiments tofindtheoptimalstructureoftheNNs.Anon-lineartansig transfer function was used in the hidden layers.One neuron was used in the output layer with a lineartransfer function, purelin. Different types of functions wereused in the trial and error procedure to find the optimalfunction that is effective on the classification success of theNN structure. Purelin was the most efficient in the outputneuron.Values of 0.25 and 0.0025 were, respectively, selectedas the coefficients of learning rate and momentum basedon trial and error. The maximum number of iterations intraining was set to 3000. The error rate convergence crite-rion for the NNs to stop learning was 0.02. NNs Toolbox inMatlab was used to implement the classifier (Howard &Meale, 1992).3.4.Error estimationError estimation was performed on classification resultsobtained from training and testing the classifiers. In evaluat-ing a classifier, half of the data was used to train the classifierwhile half of it was used for testing (Table 1). The error rate forthe NN classifier was calculated at the end of training andtesting procedures, based on the differences between targetedandcalculatedoutputs.Classificationaccuracyfortheclassifiers studied was calculated using? eN,(10)ARTICLE IN PRESSWeightsWeightsWeightsInput LayerFirst Hidden LayerSecond HiddenLayerOutputLayerErrorTargetOut12345123456784321Fig. 4 Schematic display of the MLP neural network classifier.B I O S YS T E M S E N G I N E E R I N G99 (2008) 211 219216where e is the estimate of the true error rate, N is the totalnumber of samples and e is the number of misclassified patterns.4.Results and discussionThe classification results obtained using one parametric andfivenon-parametricclassificationalgorithmsandthreedifferent feature sets are given in Tables 2 and 3.It should be noted that the successes of the classifiers arebeing compared with the classification results obtained from thehuman expert. Subjectivity is involved in the classificationperformed by the human expert even though the expertfollowed the USDA standards. Therefore, it is not expected thatan expert could perform a 100% correct classification. Possiblemeasurement errors from devices should also be considered.Also, the data used in this study were from the same varietyof apples in which the morphological properties did not differARTICLE IN PRESSTable 3 Confusion matrices for the results obtained in testingaClassifierClass4 Features5 Features9 Features123123123142220442005680Plug-in decision rule23190121013903040040310Error rate0.3220.2780.27815851604060401-Nearest neighbour2895711415703121112121Error rate0.2440.1890.24415851604060402-Nearest neighbour2895711415703121112121Error rate0.2440.1890.24416310640064003-Nearest neighbour21651174120203310400310Error rate0.2440.2440.26716130Decision tree classifierb27783130Error rate0.2441614061306220Multi-layer perceptron26133616061603022103013Error rate0.1670.1110.100aNumbers in the diagonal from the left corner to the right corner show the number of correctly classified patterns for the related class whilethe numbers off the left to the right diagonal show the number of misclassified patterns.bUsed only with four features.Table 2 Performances of the classifiers using different number of featuresClassifierClassification success, %4 Featuresa5 Featuresb9 FeaturescPlug-in decision rule67.7872.2272.221-Nearest neighbour75.5681.1175.562-Nearest neighbour75.5681.1175.563-Nearest neighbour75.5675.5673.33Decision tree classifierd75.56Multi-layer perceptron83.3388.8990.00aColour, shape defect, weight, russeting.bColour, shape defect, firmness, weight, russeting.cColour, shape defect, circumference, firmness, weight, blush percentage, russeting, size of bruises, size of natural defects.dUsed only with the sub-set including 4 features.B I O S YS T E M S E N G I N E E R I N G99 (2008) 211 219217widely enough to obtain clearly separable classes. In otherwords, a large amount of overlap existed between the groupsas can be seen in Figs. 13. This overlap may not only be dueto the physical similarities of the produce, it may also be dueto the fact that the human expert followed a somewhatsubjective classification procedure. The subjectivity of theexpert might have had an influence on the results despitethe USDA standards used, which may be explained by thedifferences in the experiences, likes and dislikes that theexpert possessed. This situation makes the problem of fruitclassification by soft computing techniques, as in this study,more complex. Therefore, in judging the results of theclassifiers, the above concerns should also be considered.Another disadvantage of the data was that the number ofpatterns per class was not equal; 128 patterns in class-1, 44patterns in class-2 and only nine patterns in class-3. Thissituation may have led to poor training performance espe-cially for class-3.Different feature sets were tested on the classifiers. Theresults obtained from these applications are reported in thefollowing sections.4.1.Effect of different feature sub-sets on classificationperformanceDifferent subsets of nine initial features were tested on theclassifiers: four, five and nine featured (Table 2) input setswere used in six different classifiers. Out of three feature setsused, the five-featured sub-set produced the most successfulclassification results in general (Table 2). Although the resultsobtained using different feature sets were close to each other,the five-featured sub-set showed better classification resultscompared to the four-featured sub-set. For the classificationapplications using K-NN classifiers (1-NN, 2-NN and 3-NN),the five-featured sub-set of features also outperformed thenine-featured sub-set, which contained all the featuresoriginallymeasured.Usingthefive-featuredsub-setoffeatures decreased the time for the classification applications,which resulted in the same or higher classification successescompared to using nine-featured subsets. Using nine featuresintheNNclassifier,however,resultedinthehighestclassification success.4.2.Comparison of classification algorithms in terms ofclassification performanceWhen the classification results were compared in terms of thesuccesses of the classification algorithms, MLP NNs yieldedthe most successful results for all the feature sets studied.This may be attributed to the ability of this classifier inassessing the non-linearities between the input features andthe output classes. The most successful result (90%, Table 2)using NNs classifier was obtained using nine features; there-fore, improved classification results can be expected withmore input features as in the case of using a nine-featuredinput set in the NNs classifier. However, MLP NNs usinga five-featured sub-set produced a classification perfor-mance (88.89%) close to that of the MLP using a nine-featuredsub-set.K-NN classifiers (1-NN, 2-NN and 3-NN) followed the MLPNN classifier in terms of classification success for all thefeature sub-sets (Table 2). The optimal feature set for thisclassifier was the five-featured input set resulting in thehighest classification success especially for 1-NN and 2-NNclassifiers. Increasing K to a value greater than three did nothelp further improve the performances of K-NN classifiers.The DT classifier, which was used only with the four-featured sub-set, performed the same as or was inferior toK-NN classifiers in general (Table 2).The statistical classifier of plug-in decision rule performedthe same (72.22%) for five- and nine- featured input sets(Table 2) while its performance was lower for the fourfeatured input set (67.78%, Table 2).Scatter plots of the first two linear discriminant variablesand the first two principal components for all observations inthe training sets for three different feature sets are shown inFigs. 13. While there is some separation of the classes usingthefirsttwodiscriminantvariables(and thefirst twoprincipal components), there are also some overlaps betweenclasses. However, there seems to be a slight improvement inthe separation of the classes on increasing the features usedfrom four to using nine. Distinction of class three was muchclearer when nine features were used instead of four or fivefeatures.4.3.Evaluation of classification results via confusionmatricesWhen the confusion matrices (Table 3) are investigated, it canbe seen that the best classifications were performed by theMLP NNs classifier. The NN classifier that used all ninefeatures provided the best classification (Table 3); only two ofthe 64 apples were misclassified for the first class, while six of22 apples were misclassified for the second class. Althoughonly four test patterns were available for class three, threetest patterns were classified correctly and the misclassifiedtest pattern was assigned to the neighbour class two. Havinga low number of samples for class three was a detrimentalsituation resulting from poor training and consequently poorperformance. The MLP NN classifier that used five inputfeatures yielded the second best classification result (Table 2)in which three of the test samples out of 64 in the first class,six out of 22 in the second class and one out of four in thethird class were misclassified (Table 3).1-NN and 2-NN classifiers used with five features providedthe best performances (Table 2) after NN classifiers. In theseclassifiers, four out of 64 apples were misclassified into theneighbour class. For the second class, on the other hand, 11apples were correctly classified while seven apples weremisclassified into class-1 and four apples were misclassifiedinto class-3 (Table 3). For the third class, two apples werecorrectly classified while two apples, one into class-1 and theother into class-2, were misclassified.For the case of using five features, 20 of the apples in class-1were misclassified into class-2 in classification with a plug- indecisionruleclassifier(Table3).Onlyoneapplewasmisclassified into class-1 out of 22 apples in class-2 whileall the apples in class-3 were misclassified into class-2.ARTICLE IN PRESSB I O S YS T E M S E N G I N E E R I N G99 (2008) 211 219218The DT classifier misclassified three of the 64 apples inclass-1 as class-2 (Table 3). Seven of the apples in class-2 weremisclassified as class-1 while another eight apples weremisclassified as class-3. Finally, all the apples in class-3 weremisclassified as class-1 and class-2.Apples in class-3 were not correctly classified at all usingPDR, 3-NN and DT classifiers. The low number of trainingpatterns for this class could be the reason for the misclassi-fications (Table 1).The same set of apples was classified using the fuzzy logictechnique in a previous study (Kavdir & Guyer, 2003) resultingin around a 89% classification success.5.ConclusionsQuality features of apples such as firmness, size, weight,shape defect, colour, blush percentage, russeting, size ofbruises and size of natural defects were measured and usedin parametric and non-parametric classifiers such as plug-indecision rule, 1-, 2- and 3-nearest neighbours, decision treeand neural networks (NNs) to determine a feasible way ofmatching feature set and the classifier for an optimalclassification result.When different feature sets were used in different classi-fiers, the feature set containing five features of colour,firmness, shape defect, weight and russeting produced theoptimum classification with less time for training (comparedto nine-featured input sets) and higher classification resultsin general, except for the NN classifier using nine features.The NN classifier using all the nine features in the input setproduced the highest classification result (90%) among all theapplications tested. The success of this
- 温馨提示:
1: 本站所有资源如无特殊说明，都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
2: 本站的文档不包含任何第三方提供的附件图纸等，如果需要附件，请联系上传者。文件的所有权益归上传用户所有。
3.本站RAR压缩包中若带图纸，网页内容里面会有图纸预览，若没有图纸预览就没有图纸。
4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
5. 人人文库网仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对用户上传分享的文档内容本身不做任何修改或编辑，并不能对任何下载内容负责。
6. 下载文件中如有侵权或不适当内容，请与我们联系，我们立即纠正。
7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
人人文库网所有资源均是用户自行上传分享，仅供网友学习交流，未经上传用户书面授权，请勿作他用。
川公网安备: 51019002004831号