Research Article A Computational Approach towards Visual...

Research ArticleA Computational Approach towards Visual Object Recognitionat Taxonomic Levels of Concepts

Zahra Sadeghi12 Babak Nadjar Araabi12 and Majid Nili Ahmadabadi12

1Cognitive Robotics Lab School of Electrical and Computer Engineering University of Tehran Tehran 14395-515 Iran2School of Cognitive Sciences Institute for Research in Fundamental Sciences (IPM) Tehran 19395-5746 Iran

Correspondence should be addressed to Zahra Sadeghi zahrasadeghiutacir and Babak Nadjar Araabi araabiutacir

Received 14 February 2015 Revised 2 June 2015 Accepted 4 June 2015

Academic Editor Thomas DeMarse

Copyright copy 2015 Zahra Sadeghi et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

It has been argued that concepts can be perceived at three main levels of abstraction Generally in a recognition system objectcategories can be viewed at three levels of taxonomic hierarchy which are known as superordinate basic and subordinate levelsFor instance ldquohorserdquo is a member of subordinate level which belongs to basic level of ldquoanimalrdquo and superordinate level of ldquonaturalobjectsrdquo Our purpose in this study is to take an investigation into visual features at each taxonomic level We first present arecognition tree which is more general in terms of inclusiveness with respect to visual representation of objects Then we focuson visual feature definition that is how objects from the same conceptual category can be visually represented at each taxonomiclevel For the first level we define global features based on frequency patterns to illustrate visual distinctions among artificial andnatural In contrast our approach for the second level is based on shape descriptors which are defined by recruiting momentbased representation Finally we show how conceptual knowledge can be utilized for visual feature definition in order to enhancerecognition of subordinate categories

1 Introduction

Categorization is the first step in recognition and so it isfundamental for perception communication and any kindof interaction with the environment [1] The goal of catego-rization is to partition the search space into different groupsin such a way that members of each group reflect a similarconcept or idea According to research in conceptual devel-opments there exist different strategies for object catego-rization Three different types of categorization are identi-fied by cognitive science researchers based on whether thesimilarities are defined by their external relations or internalpropertiesThese are known as thematic categorization scriptcategorization and taxonomic categorizationThematic cate-gorization focuses on the spatially or contiguous relationshipbetween objects For example dog and leash are in the samethematic category In script categorization objects with simi-lar roles or functionality are grouped together For instanceegg and cereal belong to the same script category Finallytaxonomic categorization refers to a hierarchy which isconstructed in an ascending order of inclusiveness (eg

like terrier-mammal-animal) [2] While the members of thefirst two types of categorization do not necessarily sharesimilar properties in the taxonomic categorization objectsare grouped based on similar observable featuresThereupontaxonomic organization is applicable to visual object catego-rization in which object appearance plays a determinant rolein object recognition and classification

This paper is organized as follows We first explain threelevels of concepts in taxonomic categorization Then inSections 3 and 4 computational models are presented forfeature definition at the first level (ie superordinate level)and second level (ie basic level) of taxonomy Finally inSection 5 we show that conceptual object representationimproves accuracy of subordinate categorization

2 Three Taxonomic Levels of Concepts

It has been well demonstrated that infants children andadults use different levels of inclusiveness for object namingand categorization [3ndash5] This semantic taxonomical model

Hindawi Publishing CorporationComputational Intelligence and NeuroscienceVolume 2015 Article ID 905421 10 pageshttpdxdoiorg1011552015905421

2 Computational Intelligence and Neuroscience

Table 1 Levels of abstraction

Level of taxonomy ExampleThe superordinate level AnimalThe basic level DogThe subordinate level Reriever

Natural Artificial

Objects

PlantAnimal Basic level

Subordinate level

Superordinate level

middot middot middotmiddot middot middot PiP1AiA1

Figure 1 Taxonomic structure of recognition used in this paper 119860119894

and 119875119894refer to the subcategories of animal and plant correspond-

ingly

known as superordinate-basic-subordinate categorization isconstituted of three levels of abstraction and is shown inTable 1 [6ndash8] The degree of inclusiveness is highest at thetop level and decreases approaching the bottom of the treeIt is not clearly known which of these levels is primarily usedin recognition of objects and it is generally accepted thatidentification of each level relies on several parameters likethe object familiarity and frequency as well as the context inwhich the object is viewed Also the number of hierarchicallevels is variable among different groups of people accordingto their level of expertise Rosch et al have shown thatpeople use the basic level as a preferred class for recognitionof objects They posit that this level constitutes optimalinformation for quick categorization [8] In contrast to thisclaim other studies have challenged this idea by showing thathuman perceive superordinate distinction prior to basic leveland it occurs at early stages of processing visual information[9ndash11] One supporting explanation behind top-down designis declared to be survival reasons because coarse informationobtained from quick processing will promote an immediateappropriate reaction [12]

In this paper we consider a more general taxonomicstructure with an onset on artificial versus natural groupsin the very first step of bifurcation of all items This is illus-trated in Figure 1 This structure is adopted according to thevisual properties of objects It has been previously shown byOliva and Torralba that scene images can be semanticallydiscriminated along artificial to natural axis at the superor-dinate level of categorization [13] The natural supercategorymight then be subdivided into animal and plant subcate-gories at the basic level Thereupon our terminology forsuperordinate basic and subordinate categories is slightlydifferent from what has been broadly used in the literature ofpsychology For instance we assume that ldquohorserdquo is locatedin subordinate level which belongs to the basic level of

ldquoanimalrdquo and the superordinate level of ldquonaturalrdquo objects Ourinvestigation for the tree structure is only devoted to naturalobjects behind which therersquos a stronger theory of hierarchicalsemantic structure For instance according to folk biologyregardless of their culture people have a similar taxonomicstructure for thinking about living subcategories as animalsand plants [14]There are also a number of studies advocatingthe superiority of tree-structure for capturing taxonomicrelationships among biological data [15 16] Following thisstructure for a classification task each image is associatedwith three different labels corresponding to each particularlevel of inclusiveness In the following sections we describeour computational approach for feature definition at eachlevel of concept For evaluation we collected benchmark datafromCaltech-101 and coil-100 image databases MPEG-7 andthe stimuli database gathered by Konkle et al [17] Moredetails about the categories and their labels are shown inTable 2 It should be mentioned that we created conceptualcategories of animal and plant based on classes available ineach database For plant categories we could only find 6 suchclasses inCaltech 101Hence in order to place an equal chancelevel of 16 in both subcategories of animals and plants weonly selected 6 subclasses for the animal class as well Thesubclasses are selected such that three of them are quadrupedanimals and the other three are birds Note that all objectsare segmented before the whole process of recognition usingannotation information associated to each object class Theobjects are then cropped to reduce the area of backgroundof images The stimuli database provided by Konkle et alcontains colorful images of isolated objects with a plain back-ground In contrast images fromMPEG-7 are isolated objectsin binary format and hence they are only used in Section 4for basic level categorization One example of each naturalsubcategory for the first two datasets is shown in Figure 2

3 Superordinate Level of Recognition

The first level of inclusiveness in hierarchy of concepts con-sists of two supplementary groups of items that is artificialand natural entities Intrinsically all objects can be consid-ered as belonging to one of the conceptual categories of eitherartificial or natural items based on their inherent source ofcreation In other words objects can be classified as eitherhuman-made (artificial) or non-human-made (natural) enti-ties Breaking up all existent items in such a way can be con-templated as the utmost general course of viewing the worldthat is objects are assumed to bemade bymankind or they arefound in the nature without human interference In additionwe are inclined to think that distinguishing objects at the firstlevel of taxonomy is independent of prior knowledge and thatthis distinction can bemade in an unsupervisedmannerThisis in accord with the top-down model proposed by Bar inwhich coarse information derived from a visual input directlyactivates similar high level representations without makingan exhaustive search to find a similar stored representationin memory [12]

The contributing role of semantic content in makinga broad distinction of images has been studied on scene

Computational Intelligence and Neuroscience 3

Table 2 Object categories in taxonomic structure

Dataset 1 Dataset 2 Dataset 3Superordinate level Natural Artificial Natural Artificial NaturalBasic level Animal Plant Animal Plant Animal Plant

Subordinate level

Flamingopigeonrooster

cougar-bodyelephantgerenuk

Sunflowerwater-lilylotus

strawberrybonsai

Joshua-tree

Obj1 obj3obj5 obj6obj7 obj8obj9 obj10obj11 obj12obj14 obj15

Birdcatdog

Bonsaigreenplant

tree

Balloonbucketcooler

horseshoemattress

Batbird

chickendeer

elephanthorse

AppleDevice 0 (flower 0)Device 1 (flower 1)Device 2 (flower 2)Device 7 (flower 3)

tree

Elephant Gerenuk FlamingoCougar-body

Bonsai

Bird Cat

Pigeon Rooster GreenplantDog

TreeBonzaiStrawberryWater-lilySunflower

Dataset 1 Dataset 2

Joshua-tree

Lotus

Figure 2 Sample of animal and plant subcategories

images [13] as well as isolated object categories [18] and fre-quency-based features have indicated efficient results in cap-turing the superordinate characteristics of objects Specifi-cally the frequency attributes of objects are defined using thefollowing equations [18]

FI = 119865 (input image) (1)

magnitude (119909 119910)

= radicRe (FI (119909 119910))2 + Im (FI (119909 119910))2(2)

phase (119909 119910) = tanminus1 (Im (FI (119909 119910))Re (FI (119909 119910))

) (3)

FreqFeat (1) = sum

119909119910isininput image

1003816100381610038161003816magnitude (119909 119910)1003816100381610038161003816 (4)

FreqFeat (2)

= sum

119909119910isininput imagelog (1+magnitude (119909 119910)) (5)

FreqFeat (3) = sum


1003816100381610038161003816phase (119909 119910)1003816100381610038161003816 (6)

where FI is the result of Fourier transform of gray scale inputimage Figure 3 illustrates the distinguishable values capturedby the visual features explained via (4) to (6) using dataset1 It can be seen that the three dimensions are all containingdistinctive values for grouping objects in two separate groupsIn addition we performed a clustering task on the obtainedfeature values to evaluate the discrimination characteristicsof the feature sets in an unsupervised mannerThe results areevaluated by using119891-measure precision recall and accuracyand are compared with Gabor [19] and C2 features [20] in


Table 3 Clustering evaluation results

Dataset 1 Dataset 2119865-measure Precision Accuracy Recall 119865-measure Precision Accuracy Recall

Frequency features 9496 9378 9500 9430 6591 6288 6212 6613Gabor feature 7180 7234 6256 7232 6260 6230 6206 6261C2 features 7591 8433 7407 7601 5127 5154 5127 5101

200 400 600 800 1000 1200 1400

1

2

3

Natural Artificial

05

15

25

35

Figure 3 Frequency features for all data In an up-down directionvertical axis represents the three dimensions defined in (4) to (6)

Table 3 In the appendix we provide further analytical figures(Figures 7 8 and 9) which indicate the potent discriminationobtained by the defined feature sets

4 Basic Level of Recognition

In this section we address the problem of basic categoryrepresentation This is the second level in the taxonomicstructure which is associated with the general classes withinnatural superordinate category The purpose of this phaseis to investigate the visual distinction between animal andplant classes Hence we deal with two broad semantic sub-categories of natural objects (ie animals and plants) Atremendous amount of research has been conducted onobject recognition based on local properties of objects (HOG[21] C2 [22] SIFT [23] and LBP [24]) In contrast in ourapproach in order to distinguish between the conceptualcategories at the basic level we utilize shape descriptorsto extract global main discriminations between animal andplant categories The theory behind this approach is that thecategories of animal and plant are distinguishable in form andconfiguration and hence using global features by applyingshape descriptors can be profitable

41 Method For modeling object shapes we employ momentdescriptors to quantitatively capture the principal shapeinformation of objects To this end image binarization is car-ried out on all images Samples of resultant images after bina-rization are shown in Figure 4 This process removes texturaldetails but preserves the whole shape of objects Therefore

only holistic representation of objects is taken into consid-eration Note that the binarization process is performed inorder to provide global outline of objects In essence in thissection we are looking for computational evidence to supportpsychological preference for basic level categorization as theentry level It has been proposed that low spatial frequencyinformation which forms the global appearance of objects isperceived before fine properties [12] Our results provide sup-port by demonstrating that basic categorization is not reliedon in local processing and by employing global informationthrough a shape based approach we can still reach highdistinction between broad categories defined at this level

As we mentioned before the proposed feature vectorsare constructed by moment-based descriptors To this endwe computed the first eight standardized moments as well asZernikemomentsThe simplestmoment computes the centerof mass in both directionsThe secondmoment measures thevariation from the center of the object in vertical andhorizon-tal directions Skewness is the third moment which measuresthe orientation of a distribution in the 119909 and 119910 directionsWe therefore used the absolute value of this parameter totreat equally the left and right skewness Fourth moment iskurtosis and deals with the peakedness and tail weight ofa distribution The fifth to eighth moments quantify highershape parameters We further calculated Zernike momentsto obtain richer shape characteristics of objects Zernikemoments are constructed by projection over a sequence oforthogonal basis polynomials [25] and they have shown tobe effective in shape classification tasks [26 27] In ourexperiments we used the magnitude of Zernike momentsover 20 basis functions of order 6 (we used public codesreleased byChristianWolf available at httpliriscnrsfrchris-tianwolf) We then concatenated the feature vector obtainedfrom standard moments with Zernike moments resulting ina 32-dimensional feature vector

bsFeat (1 16) = [1205831 1205838]

bsFeat (16 32) = [10038161003816100381610038161198601198991198981003816100381610038161003816] 119899 = 6 119898 = 0

(7)

where 120583119899is a two-dimensional vector of the 119899th-order mo-

ment of the input image on both 119909 and 119910 directions and 119860119899119898

is the projection of the image into Zernike basis function oforder 119899 with repetition119898

42 Results and Discussion To highlight the efficacy of thismethod we compared our results with C2 features [22] andHOG descriptors [21] which are known as successful tech-niques for object recognition In all cases SVM classifierswith linear kernels are used The C2 features are computedby HMAX model in a four-layered architecture (two S layers


Table 4 Comparison results of classification on basic conceptual categories Results are averaged over 10 iterations Time complexity isaveraged over all train samples

Features

Dataset 1 Dataset 3Feature vectordimensions

Accuracy Averageprocessing time

per sample


per sampleAnimalclass

Plantclass

Total (over alltest samples)

Animalclass

Plantclass


C2 8655(266)

8433(298)

8527(241) 356 9350

(228)9550(314)

9450(148) 723 200

HOG 8237(321)

8033(162)

8113(157) 00385 8566

(402)9466(258)

9016(203) 01452 128

Moment-basedmethod

8667(257)

8535(278)

8571(79) 01465 9466

(316)9516(199)

9492(193) 02902 32

Gerenuk Horse Elephant

Device 0 (flower 0) Apple

Dataset 3Dataset 1

Rooster

Water-lily Joshua-tree Device 0 (flower 0) AppleWater-lily Joshua-tree

Figure 4 Samples of binary images of objects

and two C layers which perform template matching andmax-pooling resp) The final C2 features are the result ofresemblance to the stored local patches (in our case 200patches) Histogram of Gradients (HOG) descriptor is cre-ated by counting occurrence of different orientations insidegrids and concatenating them into a vector We applied thebasic form of HOG algorithm by dividing each image into4 by 4 nonoverlapping blocks and calculating orientationhistograms with 8 bins over each block Thereupon eachinput image is described with 200 dimensions using HMAXmodel and with 128 dimensions using HOGmethod In con-trast in the proposedmoment-based approach each image isrepresented with an input vector of length 32 Neverthelessit can be understood from Table 4 that the proposed global

approach achieves better performance compared to the otherlocal powerfulmethods (higher total accuracy and lower timecomplexity in comparison to C2 features) It is remarkablethat while the set of statistical moments are simple computa-tionally cheap and fast to be processed on both training andtest phases they attain high performance It can be arguedthat shape-based approaches are preferable in situationswhenhigh resolution images are not available or cannot be storeddue to memory space issues and information bottlenecksThis may not seem to be a serious problem regarding thetremendous development inmemory technologies Howeverit is highly profitable and biologically arguable to take anapproachwhich is not dependent on consuming large volumeof memory The results also suggest that shape properties


are a rich source of information for classification of generalcategories of animal and plant One explaining factor is highdegree of feature sharedness amongmembers of general con-cepts [28] which boosts the structural similarity within eachgroup Studies towards global representation of objects arealso important for mind and brain research For instanceit has been shown that patients with semantic impairmenthave difficulties to access subordinate knowledge [29] but yetnot much is known about the characteristics of the type ofknowledge and its internal representation in brainThe lateraloccipital complex (LOC) in human brain has been foundto be involved in visual shape processing of objects [30] Inparticular it has been shown that LOC activation is relatedto shape characteristic of objects rather than specific featuressuch as edge [31] More studies and experiments are requiredto be conducted in order to probe the interplay between lowlevel visual area (eg V1) and higher level visual area (egLOC) as well as the underlying visual mechanism regardingto relationship between local and global visual processing

5 Subordinate Level of Recognition

While the categories associated to the superordinate and basicconcepts are demonstrated to be well distinguishable byutilizing global features (Sections 3 and 4) detailed informa-tion is required in order to capture fine distinction withinsubordinate categories For example while quadruped ani-mals such as cougar and elephant can be distinguished fromflowers such as sunflower and water lily by using global shapeinformation further local processing is required to tell themapart Initial support for this argument comes from biologicalstudies about coarse-to-fine processing in visual system anal-ysis [32 33] or global-to-local approaches [34ndash36] In this sec-tion we investigate whether visual characteristics collectedfrom conceptual space encompass efficient information forrecognition of subcategory objects In other words we ques-tion whether it would be beneficial to define feature vectorsfor subcategories of animal and plant by driving specificinformation about each conceptual space In essence we pro-pose an approach that utilizes conceptual space informationfor feature extraction For this purpose we divide naturalsuperordinate category into two subcategories of animal andplant on training samples Then we develop local featuresbased on information extracted from each subspace

51 Method Our strategies are developed based on the ideathat conceptual knowledge can provide detailed informationabout structure of each basic class To peruse this idea wetake an approach similar to prototype matching in which weuse PCA method All images are first cropped to eliminateborder area and then rescaled to 100 times 100 pixels Next theeigenvectors of covariance matrix of all training images aregenerated Note that instead of generating the covariancematrix of the stimuli set (ie 119878119878119879) which is a very large(119873

2times 119873

2 119873 = 100) dimensional matrix we compute

the covariance matrix associated to the transpose of 119878 (ie119878119879119878) [37] Thus the relationship between the eigenvectors of

covariance matrix of 119878 (ie 119906119894) and the eigenvectors of the

covariance of 119878119879 (ie V119894) can be expressed by

119906119894= 119878V119894 (8)

Feature vectors corresponding to each image are generatedby projection of imagersquos pixels on the computed eigenvectorsWe create eigenvectors in two modes (1) flat mode in whicheigenvectors are calculated over all images of animals andplants (flat space) and (2) conceptual mode in which eigen-vectors are calculated on each subspace of animal and plantseparately and then the results are concatenated In otherwords in flat recognition feature representation for imagesis made by projection on the eigenvectors derived fromthe flat space whereas in the conceptual recognition thefeature representation of each image is built by concatenatingthe projection on both subspaces The following describessubcategory feature creation

subFeat (119878) =

119906119879

119860119875119878 flat mode

[119906119879

119860119878 119906119879

119875119878] conceptual mode

(9)

where 119906119860119875

denotes the eigenvectors obtained from trainingsamples including both animal and plant categories while 119906

119860

and 119906119875represent eigenvectors computed over animal samples

and plant samples respectivelyFigure 5 demonstrates the first eight eigenvectors associ-

ated to each conceptual subspace using dataset 1 It can beinferred that eigenvectors associated to the flat space are notas informative as eigenvectors derived from conceptual space

52 Results and Discussion For evaluation we comparedaccuracy of flat recognition to that of conceptual recognitionIn our method different random sets of train and test sam-ples are created at each iteration and the results are averagedover ten independent runs In addition we use SVM clas-sifier with linear kernel for evaluation The performanceassociated to subordinate categories of dataset 1 in terms oftheir mean and standard deviation of the mean is illustratedin Figure 6 The number of training samples used in eachexperiment is indicated by nt variable Note that total numberof eigenvectors is equal to the number of training imagesused in each subspace This is denoted by total eigenvectorsIt can be implied that the classification performance hasbeen improved in almost all classes We also evaluated thepercentage improvements through all categories To this endthe number of training samples is changed between 50 and75 of whole data and the feature dimensions are tested withboth total number of obtained eigenvectors and half of themThe best obtained performance of the average percentageimprovement over all classes is 2212 for dataset 1 and3026 for dataset 2 These results suggest that deployingknowledge of conceptual categories of animal and plantimproves the accuracy of subcategory recognition Theseresults are achieved due to high intraclass similarity of basicgroups and their low interclass similarity The abstract infor-mation that accounted for the main characteristic of featuresof each conceptual space is captured by the aid of eigenvectors


(a)

(b)

(c)

Figure 5 Eigen matrices associated to (a) flat space (b) conceptual animal subspace and (c) conceptual plant subspace

derived from covariancematrix corresponding to each spaceTherefore projection of images on both conceptual spacesconcretely measures the similarity proportion of each spaceThe weight vectors obtained after image projection are thenserved as an initial prediction of subcategory candidates The

ultimate decision is made by a classification task that utilizesa vector of predictions In contrast in the flat mode whereno specific conceptual knowledge is provided eigenvectorscarry combined information from both spaces and hencethey fail to attain fine subcategory distinctions Note that in


0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 140 2 4 6 8 10 12 14

10

20

30

40

0

20

40

60

0

20

40

60

0

20

40

60

10

20

30

40

10

20

30

40

0

20

40

60

Stra

wbe

rry

0

20

40

60

FlatConceptual

FlatConceptual

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk

Stra

wbe

rry

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk

Total eigenvectors Half eigenvectorsn

t=24

nt=

20

nt=

30

nt=

18

nt=

24

nt=

20

nt=

30

nt=

18

Figure 6 Categorization accuracy (nt number of training samples) Total number of eigenvectors are equal to the total number of trainingsamples

this section we showed only one possible implementationthat reveals the impact of conceptual knowledge on boostingthe object recognition rate Our future work will concentrateon developing more powerful methods that benefit from tax-onomic knowledge

6 Conclusion

In this study we investigated visual representation of con-cepts at three levels of inclusiveness The concepts of eachlevel are known as superordinate basic and subordinate cate-gories Tomake distinction between superordinate categoriesat the first level (ie artificial and natural concepts) weused energy of frequency spectrum of images and showedits superiority compared to two other methods For basiccategory representation in the second level (ie animal andplant concepts) we proposed to utilize moment descriptorsin order to capture the differences in shape rather than localpatches of images The results demonstrated overall better

performance to that of local based methods Finally weshowed that space decomposition based on conceptual cate-gories can be beneficial in terms of accuracy in recognitionof subordinate object classes Our attempt in all the threephases was motivated from cognitive theories to delineate aconsistent computational model The superordinate and sub-ordinate categories both stand at a lower level of cue validitythan basic level which indicates the basic category is the mostinclusive level [8] Accordingly in our approach the first andthird levels of recognition rely on gray scale information butthe second level of recognition is based on shape propertiesobtained through processing of binary images

Appendix

The corresponding scatter plot associated to the two firstdimensions of frequency features on dataset 1 are plotted inFigure 7 Natural and artificial objects are denoted by redcircles and green dots consecutivelyThis figure indicates that


01

23

4

45

67

814

15

16

145

155

165

NaturalArtificial

times104

times104

times107

Figure 7 Scatter plot for all data represented by frequency featuresArtificial and natural images are represented with green and redcircles correspondingly

06

07

08

09

055

065

075

085

095

Frequency features

C2 features

1

0605 07 08 09055 065 075 085 095

Reca

ll

Precision

Gabor features

Figure 8 Precision-recall curve The results are obtained by usingdifferent threshold values on the result of fuzzy clustering

features defined at the superordinate level map data into aspace where they are linearly separable The precision-recallcurve that is obtained by using different threshold values onthe result of fuzzy c-means clustering is shown in Figure 8Besides fuzzy memberships for all artificial and natural testdata are illustrated in Figure 9 The blue bars indicate thedegree of naturalness

0 200 400 600 800 1000 1200 1400 16000

1

Fuzz

y m

embe

rshi

p

ArtificialNatural

010203040506070809

Figure 9 Fuzzy membership grade Each bar shows the degreemembership of each data to natural fuzzy cluster

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] T Frey M Gelhausen and G Saake ldquoCategorization of con-cerns a categorical program comprehension modelrdquo in Pro-ceedings of the 3rd ACM SIGPLAN Workshop on Evaluationand Usability of Programming Languages and Tools pp 73ndash82October 2011

[2] S P Nguyen and G L Murphy ldquoAn apple is more than just afruit cross-classification in childrenrsquos conceptsrdquo Child Develop-ment vol 74 no 6 pp 1783ndash1806 2003

[3] J Liu R M Golinkoff and K Sak ldquoOne cow does not ananimal make young children can extend novel words at thesuperordinate levelrdquoChild Development vol 72 no 6 pp 1674ndash1694 2001

[4] A E Ellis and LMOakes ldquoInfants flexibly use different dimen-sions to categorize objectsrdquo Developmental Psychology vol 42no 6 pp 1000ndash1011 2006

[5] M H Bornstein and M E Arterberry ldquoThe development ofobject categorization in young children hierarchical inclusive-ness age perceptual attribute and group versus individual anal-ysesrdquo Developmental Psychology vol 46 no 2 pp 350ndash3652010

[6] C B Mervis and E Rosch ldquoCategorization of natural objectsrdquoAnnual Review of Psychology vol 32 no 1 pp 89ndash115 1981

[7] E H Rosch ldquoNatural categoriesrdquo Cognitive Psychology vol 4no 3 pp 328ndash350 1973

[8] E Rosch C B Mervis W D Gray D M Johnson and PBoyes-Braem ldquoBasic objects in natural categoriesrdquo CognitivePsychology vol 8 no 3 pp 382ndash439 1976

[9] M Praszlig C Grimsen M Konig and M Fahle ldquoUltra rapidobject categorization effects of level animacy and contextrdquoPLoS ONE vol 8 no 6 Article ID e68051 2013

[10] M R Greene and A Oliva ldquoThe briefest of glances the timecourse of natural scene understandingrdquo Psychological Sciencevol 20 no 4 pp 464ndash472 2009


[11] L C Loschky and A M Larson ldquoThe naturalman-madedistinction is made before basic-level distinctions in scene gistprocessingrdquo Visual Cognition vol 18 no 4 pp 513ndash536 2010

[12] M Bar ldquoA cortical mechanism for triggering top-down facili-tation in visual object recognitionrdquo Journal of Cognitive Neuro-science vol 15 no 4 pp 600ndash609 2003

[13] A Oliva and A Torralba ldquoModeling the shape of the scenea holistic representation of the spatial enveloperdquo InternationalJournal of Computer Vision vol 42 no 3 pp 145ndash175 2001

[14] S Atran ldquoFolk biology and the anthropology of science cog-nitive universals and cultural particularsrdquo Behavioral and BrainSciences vol 21 no 4 pp 547ndash609 1998

[15] C Kemp and J B Tenenbaum ldquoStructured statistical modelsof inductive reasoningrdquo Psychological Review vol 116 no 1 pp20ndash58 2009

[16] C Kemp and J B Tenenbaum ldquoThe discovery of structuralformrdquo Proceedings of the National Academy of Sciences of theUnited States of America vol 105 no 31 pp 10687ndash10692 2008

[17] T Konkle T F Brady G A Alvarez and A Oliva ldquoConceptualdistinctiveness supports detailed visual long-term memory forreal-world objectsrdquo Journal of Experimental Psychology Gen-eral vol 139 no 3 pp 558ndash578 2010

[18] Z Sadeghi M N Ahmadabadi and B N Araabi ldquoUnsuper-vised categorization of objects into artificial and natural super-ordinate classes using features from low-level visionrdquo Interna-tional Journal of Image Processing vol 7 no 4 pp 314ndash429 2013

[19] M Kim C Park and K Koo ldquoNaturalman-made object clas-sification based on gabor characteristicsrdquo in Image and VideoRetrieval vol 3568 of Lecture Notes in Computer Science pp550ndash559 Springer Berlin Germany 2005

[20] M Riesenhuber and T Poggio ldquoHierarchical models of objectrecognition in cortexrdquo Nature Neuroscience vol 2 no 11 pp1019ndash1025 1999

[21] N Dalal and B Triggs ldquoHistograms of oriented gradients forhuman detectionrdquo in Proceedings of the IEEE Computer SocietyConference on Computer Vision and Pattern Recognition (CVPRrsquo05) vol 1 pp 886ndash893 June 2005

[22] T Serre L Wolf S Bileschi M Riesenhuber and T PoggioldquoRobust object recognition with cortex-like mechanismsrdquo IEEETransactions on Pattern Analysis and Machine Intelligence vol29 no 3 pp 411ndash426 2007

[23] D G Lowe ldquoObject recognition from local scale-invariantfeaturesrdquo inProceedings of the 7th IEEE International Conferenceon Computer Vision (ICCV rsquo99) vol 2 pp 1150ndash1157 September1999

[24] T Ojala M Pietikainen and T Maenpaa ldquoMultiresolutiongray-scale and rotation invariant texture classificationwith localbinary patternsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 24 no 7 pp 971ndash987 2002

[25] A Khotanzad and Y H Hong ldquoInvariant image recognition byZernike momentsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 12 no 5 pp 489ndash497 1990

[26] Y Ying J Gui and X Rao ldquoFruit shape classification basedon Zernike momentsrdquo Journal of Jiangsu University (NaturalScience Edition) vol 28 no 1 pp 1ndash3 2007

[27] T Arif Z Shaaban L Krekor and S Baba ldquoObject classificationvia geometrical zernike and legendre momentsrdquo Journal ofTheoretical and Applied Information Technology vol 7 no 1 pp31ndash37 2009

[28] J F Marques ldquoThe generalspecific breakdown of semanticmemory and the nature of superordinate knowledge insights

from superordinate and basic-level feature normsrdquo CognitiveNeuropsychology vol 24 no 8 pp 879ndash903 2007

[29] E K Warrington ldquoThe selective impairment of semanticmemoryrdquoTheQuarterly Journal of Experimental Psychology vol27 no 4 pp 635ndash657 1975

[30] K Grill-Spector Z Kourtzi and N Kanwisher ldquoThe lateraloccipital complex and its role in object recognitionrdquo VisionResearch vol 41 no 10-11 pp 1409ndash1422 2001

[31] Z Kourtzi and N Kanwisher ldquoRepresentation of perceivedobject shape by the human lateral occipital complexrdquo Sciencevol 293 no 5534 pp 1506ndash1509 2001

[32] D Marr Vision A Computational Investigation into the HumanRepresentation and Processing of Visual Information HenryHoltCompany Inc New York NY USA 1982

[33] P G Schyns and A Oliva ldquoFrom blobs to boundary edges evi-dence for time-and spatial-scale-dependent scene recognitionrdquoPsychological Science vol 5 no 4 pp 195ndash200 1994

[34] R Kimchi ldquoPrimacy of wholistic processing and globallocalparadigm a critical reviewrdquo Psychological Bulletin vol 112 no1 pp 24ndash38 1992

[35] D Navon ldquoForest before trees the precedence of global featuresin visual perceptionrdquoCognitive Psychology vol 9 no 3 pp 353ndash383 1977

[36] B C Love J N Rouder and E J Wisniewski ldquoA structuralaccount of global and local processingrdquo Cognitive Psychologyvol 38 no 2 pp 291ndash316 1999

[37] M Turk and A Pentland ldquoEigenfaces for recognitionrdquo Journalof Cognitive Neuroscience vol 3 no 1 pp 71ndash86 1991

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



Table 1 Levels of abstraction

Level of taxonomy ExampleThe superordinate level AnimalThe basic level DogThe subordinate level Reriever

Natural Artificial

Objects

PlantAnimal Basic level

Subordinate level

Superordinate level

middot middot middotmiddot middot middot PiP1AiA1

Figure 1 Taxonomic structure of recognition used in this paper 119860119894

and 119875119894refer to the subcategories of animal and plant correspond-

ingly

known as superordinate-basic-subordinate categorization isconstituted of three levels of abstraction and is shown inTable 1 [6ndash8] The degree of inclusiveness is highest at thetop level and decreases approaching the bottom of the treeIt is not clearly known which of these levels is primarily usedin recognition of objects and it is generally accepted thatidentification of each level relies on several parameters likethe object familiarity and frequency as well as the context inwhich the object is viewed Also the number of hierarchicallevels is variable among different groups of people accordingto their level of expertise Rosch et al have shown thatpeople use the basic level as a preferred class for recognitionof objects They posit that this level constitutes optimalinformation for quick categorization [8] In contrast to thisclaim other studies have challenged this idea by showing thathuman perceive superordinate distinction prior to basic leveland it occurs at early stages of processing visual information[9ndash11] One supporting explanation behind top-down designis declared to be survival reasons because coarse informationobtained from quick processing will promote an immediateappropriate reaction [12]

In this paper we consider a more general taxonomicstructure with an onset on artificial versus natural groupsin the very first step of bifurcation of all items This is illus-trated in Figure 1 This structure is adopted according to thevisual properties of objects It has been previously shown byOliva and Torralba that scene images can be semanticallydiscriminated along artificial to natural axis at the superor-dinate level of categorization [13] The natural supercategorymight then be subdivided into animal and plant subcate-gories at the basic level Thereupon our terminology forsuperordinate basic and subordinate categories is slightlydifferent from what has been broadly used in the literature ofpsychology For instance we assume that ldquohorserdquo is locatedin subordinate level which belongs to the basic level of

ldquoanimalrdquo and the superordinate level of ldquonaturalrdquo objects Ourinvestigation for the tree structure is only devoted to naturalobjects behind which therersquos a stronger theory of hierarchicalsemantic structure For instance according to folk biologyregardless of their culture people have a similar taxonomicstructure for thinking about living subcategories as animalsand plants [14]There are also a number of studies advocatingthe superiority of tree-structure for capturing taxonomicrelationships among biological data [15 16] Following thisstructure for a classification task each image is associatedwith three different labels corresponding to each particularlevel of inclusiveness In the following sections we describeour computational approach for feature definition at eachlevel of concept For evaluation we collected benchmark datafromCaltech-101 and coil-100 image databases MPEG-7 andthe stimuli database gathered by Konkle et al [17] Moredetails about the categories and their labels are shown inTable 2 It should be mentioned that we created conceptualcategories of animal and plant based on classes available ineach database For plant categories we could only find 6 suchclasses inCaltech 101Hence in order to place an equal chancelevel of 16 in both subcategories of animals and plants weonly selected 6 subclasses for the animal class as well Thesubclasses are selected such that three of them are quadrupedanimals and the other three are birds Note that all objectsare segmented before the whole process of recognition usingannotation information associated to each object class Theobjects are then cropped to reduce the area of backgroundof images The stimuli database provided by Konkle et alcontains colorful images of isolated objects with a plain back-ground In contrast images fromMPEG-7 are isolated objectsin binary format and hence they are only used in Section 4for basic level categorization One example of each naturalsubcategory for the first two datasets is shown in Figure 2

3 Superordinate Level of Recognition

The first level of inclusiveness in hierarchy of concepts con-sists of two supplementary groups of items that is artificialand natural entities Intrinsically all objects can be consid-ered as belonging to one of the conceptual categories of eitherartificial or natural items based on their inherent source ofcreation In other words objects can be classified as eitherhuman-made (artificial) or non-human-made (natural) enti-ties Breaking up all existent items in such a way can be con-templated as the utmost general course of viewing the worldthat is objects are assumed to bemade bymankind or they arefound in the nature without human interference In additionwe are inclined to think that distinguishing objects at the firstlevel of taxonomy is independent of prior knowledge and thatthis distinction can bemade in an unsupervisedmannerThisis in accord with the top-down model proposed by Bar inwhich coarse information derived from a visual input directlyactivates similar high level representations without makingan exhaustive search to find a similar stored representationin memory [12]

The contributing role of semantic content in makinga broad distinction of images has been studied on scene




Subordinate level




strawberrybonsai

Joshua-tree


Birdcatdog

Bonsaigreenplant

tree

Balloonbucketcooler

horseshoemattress

Batbird

chickendeer

elephanthorse


tree


Bonsai

Bird Cat



Dataset 1 Dataset 2

Joshua-tree

Lotus




magnitude (119909 119910)

= radicRe (FI (119909 119910))2 + Im (FI (119909 119910))2(2)


) (3)

FreqFeat (1) = sum


1003816100381610038161003816magnitude (119909 119910)1003816100381610038161003816 (4)

FreqFeat (2)

= sum


FreqFeat (3) = sum


1003816100381610038161003816phase (119909 119910)1003816100381610038161003816 (6)






200 400 600 800 1000 1200 1400

1

2

3

Natural Artificial

05

15

25

35








bsFeat (1 16) = [1205831 1205838]

bsFeat (16 32) = [10038161003816100381610038161198601198991198981003816100381610038161003816] 119899 = 6 119898 = 0

(7)







Features



per sample



Plantclass


Animalclass

Plantclass


C2 8655(266)

8433(298)

8527(241) 356 9350

(228)9550(314)

9450(148) 723 200

HOG 8237(321)

8033(162)

8113(157) 00385 8566

(402)9466(258)

9016(203) 01452 128

Moment-basedmethod

8667(257)

8535(278)

8571(79) 01465 9466

(316)9516(199)

9492(193) 02902 32



Dataset 3Dataset 1

Rooster










2times 119873





119906119894= 119878V119894 (8)


subFeat (119878) =

119906119879

119860119875119878 flat mode

[119906119879

119860119878 119906119879


(9)

where 119906119860119875


119860






(a)

(b)

(c)





0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 140 2 4 6 8 10 12 14

10

20

30

40

0

20

40

60

0

20

40

60

0

20

40

60

10

20

30

40

10

20

30

40

0

20

40

60

Stra

wbe

rry

0

20

40

60

FlatConceptual

FlatConceptual

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk

Stra

wbe

rry

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk


t=24

nt=

20

nt=

30

nt=

18

nt=

24

nt=

20

nt=

30

nt=

18



6 Conclusion



Appendix



01

23

4

45

67

814

15

16

145

155

165

NaturalArtificial

times104

times104

times107


06

07

08

09

055

065

075

085

095

Frequency features

C2 features

1

0605 07 08 09055 065 075 085 095

Reca

ll

Precision

Gabor features



0 200 400 600 800 1000 1200 1400 16000

1

Fuzz

y m

embe

rshi

p

ArtificialNatural

010203040506070809




References















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






Subordinate level




strawberrybonsai

Joshua-tree


Birdcatdog

Bonsaigreenplant

tree

Balloonbucketcooler

horseshoemattress

Batbird

chickendeer

elephanthorse


tree


Bonsai

Bird Cat



Dataset 1 Dataset 2

Joshua-tree

Lotus




magnitude (119909 119910)

= radicRe (FI (119909 119910))2 + Im (FI (119909 119910))2(2)


) (3)

FreqFeat (1) = sum


1003816100381610038161003816magnitude (119909 119910)1003816100381610038161003816 (4)

FreqFeat (2)

= sum


FreqFeat (3) = sum


1003816100381610038161003816phase (119909 119910)1003816100381610038161003816 (6)






200 400 600 800 1000 1200 1400

1

2

3

Natural Artificial

05

15

25

35








bsFeat (1 16) = [1205831 1205838]

bsFeat (16 32) = [10038161003816100381610038161198601198991198981003816100381610038161003816] 119899 = 6 119898 = 0

(7)







Features



per sample



Plantclass


Animalclass

Plantclass


C2 8655(266)

8433(298)

8527(241) 356 9350

(228)9550(314)

9450(148) 723 200

HOG 8237(321)

8033(162)

8113(157) 00385 8566

(402)9466(258)

9016(203) 01452 128

Moment-basedmethod

8667(257)

8535(278)

8571(79) 01465 9466

(316)9516(199)

9492(193) 02902 32



Dataset 3Dataset 1

Rooster










2times 119873





119906119894= 119878V119894 (8)


subFeat (119878) =

119906119879

119860119875119878 flat mode

[119906119879

119860119878 119906119879


(9)

where 119906119860119875


119860






(a)

(b)

(c)





0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 140 2 4 6 8 10 12 14

10

20

30

40

0

20

40

60

0

20

40

60

0

20

40

60

10

20

30

40

10

20

30

40

0

20

40

60

Stra

wbe

rry

0

20

40

60

FlatConceptual

FlatConceptual

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk

Stra

wbe

rry

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk


t=24

nt=

20

nt=

30

nt=

18

nt=

24

nt=

20

nt=

30

nt=

18



6 Conclusion



Appendix



01

23

4

45

67

814

15

16

145

155

165

NaturalArtificial

times104

times104

times107


06

07

08

09

055

065

075

085

095

Frequency features

C2 features

1

0605 07 08 09055 065 075 085 095

Reca

ll

Precision

Gabor features



0 200 400 600 800 1000 1200 1400 16000

1

Fuzz

y m

embe

rshi

p

ArtificialNatural

010203040506070809




References















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in







200 400 600 800 1000 1200 1400

1

2

3

Natural Artificial

05

15

25

35








bsFeat (1 16) = [1205831 1205838]

bsFeat (16 32) = [10038161003816100381610038161198601198991198981003816100381610038161003816] 119899 = 6 119898 = 0

(7)







Features



per sample



Plantclass


Animalclass

Plantclass


C2 8655(266)

8433(298)

8527(241) 356 9350

(228)9550(314)

9450(148) 723 200

HOG 8237(321)

8033(162)

8113(157) 00385 8566

(402)9466(258)

9016(203) 01452 128

Moment-basedmethod

8667(257)

8535(278)

8571(79) 01465 9466

(316)9516(199)

9492(193) 02902 32



Dataset 3Dataset 1

Rooster










2times 119873





119906119894= 119878V119894 (8)


subFeat (119878) =

119906119879

119860119875119878 flat mode

[119906119879

119860119878 119906119879


(9)

where 119906119860119875


119860






(a)

(b)

(c)





0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 140 2 4 6 8 10 12 14

10

20

30

40

0

20

40

60

0

20

40

60

0

20

40

60

10

20

30

40

10

20

30

40

0

20

40

60

Stra

wbe

rry

0

20

40

60

FlatConceptual

FlatConceptual

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk

Stra

wbe

rry

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk


t=24

nt=

20

nt=

30

nt=

18

nt=

24

nt=

20

nt=

30

nt=

18



6 Conclusion



Appendix



01

23

4

45

67

814

15

16

145

155

165

NaturalArtificial

times104

times104

times107


06

07

08

09

055

065

075

085

095

Frequency features

C2 features

1

0605 07 08 09055 065 075 085 095

Reca

ll

Precision

Gabor features



0 200 400 600 800 1000 1200 1400 16000

1

Fuzz

y m

embe

rshi

p

ArtificialNatural

010203040506070809




References















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





Features



per sample



Plantclass


Animalclass

Plantclass


C2 8655(266)

8433(298)

8527(241) 356 9350

(228)9550(314)

9450(148) 723 200

HOG 8237(321)

8033(162)

8113(157) 00385 8566

(402)9466(258)

9016(203) 01452 128

Moment-basedmethod

8667(257)

8535(278)

8571(79) 01465 9466

(316)9516(199)

9492(193) 02902 32



Dataset 3Dataset 1

Rooster










2times 119873





119906119894= 119878V119894 (8)


subFeat (119878) =

119906119879

119860119875119878 flat mode

[119906119879

119860119878 119906119879


(9)

where 119906119860119875


119860






(a)

(b)

(c)





0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 140 2 4 6 8 10 12 14

10

20

30

40

0

20

40

60

0

20

40

60

0

20

40

60

10

20

30

40

10

20

30

40

0

20

40

60

Stra

wbe

rry

0

20

40

60

FlatConceptual

FlatConceptual

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk

Stra

wbe

rry

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk


t=24

nt=

20

nt=

30

nt=

18

nt=

24

nt=

20

nt=

30

nt=

18



6 Conclusion



Appendix



01

23

4

45

67

814

15

16

145

155

165

NaturalArtificial

times104

times104

times107


06

07

08

09

055

065

075

085

095

Frequency features

C2 features

1

0605 07 08 09055 065 075 085 095

Reca

ll

Precision

Gabor features



0 200 400 600 800 1000 1200 1400 16000

1

Fuzz

y m

embe

rshi

p

ArtificialNatural

010203040506070809




References















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in








2times 119873





119906119894= 119878V119894 (8)


subFeat (119878) =

119906119879

119860119875119878 flat mode

[119906119879

119860119878 119906119879


(9)

where 119906119860119875


119860






(a)

(b)

(c)





0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 140 2 4 6 8 10 12 14

10

20

30

40

0

20

40

60

0

20

40

60

0

20

40

60

10

20

30

40

10

20

30

40

0

20

40

60

Stra

wbe

rry

0

20

40

60

FlatConceptual

FlatConceptual

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk

Stra

wbe

rry

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk


t=24

nt=

20

nt=

30

nt=

18

nt=

24

nt=

20

nt=

30

nt=

18



6 Conclusion



Appendix



01

23

4

45

67

814

15

16

145

155

165

NaturalArtificial

times104

times104

times107


06

07

08

09

055

065

075

085

095

Frequency features

C2 features

1

0605 07 08 09055 065 075 085 095

Reca

ll

Precision

Gabor features



0 200 400 600 800 1000 1200 1400 16000

1

Fuzz

y m

embe

rshi

p

ArtificialNatural

010203040506070809




References















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




(a)

(b)

(c)





0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 140 2 4 6 8 10 12 14

10

20

30

40

0

20

40

60

0

20

40

60

0

20

40

60

10

20

30

40

10

20

30

40

0

20

40

60

Stra

wbe

rry

0

20

40

60

FlatConceptual

FlatConceptual

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk

Stra

wbe

rry

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk


t=24

nt=

20

nt=

30

nt=

18

nt=

24

nt=

20

nt=

30

nt=

18



6 Conclusion



Appendix



01

23

4

45

67

814

15

16

145

155

165

NaturalArtificial

times104

times104

times107


06

07

08

09

055

065

075

085

095

Frequency features

C2 features

1

0605 07 08 09055 065 075 085 095

Reca

ll

Precision

Gabor features



0 200 400 600 800 1000 1200 1400 16000

1

Fuzz

y m

embe

rshi

p

ArtificialNatural

010203040506070809




References















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 14

0 2 4 6 8 10 12 140 2 4 6 8 10 12 14

10

20

30

40

0

20

40

60

0

20

40

60

0

20

40

60

10

20

30

40

10

20

30

40

0

20

40

60

Stra

wbe

rry

0

20

40

60

FlatConceptual

FlatConceptual

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk

Stra

wbe

rry

Elep

hant

Flam

ingo

Cou

gar-

body

Bons

ai

Pige

on

Roos

ter

Wat

er-li

ly

Sunfl

ower

Josh

ua-tr

ee

Lotu

s

Ger

enuk


t=24

nt=

20

nt=

30

nt=

18

nt=

24

nt=

20

nt=

30

nt=

18



6 Conclusion



Appendix



01

23

4

45

67

814

15

16

145

155

165

NaturalArtificial

times104

times104

times107


06

07

08

09

055

065

075

085

095

Frequency features

C2 features

1

0605 07 08 09055 065 075 085 095

Reca

ll

Precision

Gabor features



0 200 400 600 800 1000 1200 1400 16000

1

Fuzz

y m

embe

rshi

p

ArtificialNatural

010203040506070809




References















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




01

23

4

45

67

814

15

16

145

155

165

NaturalArtificial

times104

times104

times107


06

07

08

09

055

065

075

085

095

Frequency features

C2 features

1

0605 07 08 09055 065 075 085 095

Reca

ll

Precision

Gabor features



0 200 400 600 800 1000 1200 1400 16000

1

Fuzz

y m

embe

rshi

p

ArtificialNatural

010203040506070809




References















































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in







































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



Research Article A Computational Approach towards Visual...

Documents

Transcript of Research Article A Computational Approach towards Visual...