Research Article Completed Local Ternary Pattern for ...downloads.hindawi.com/journals/tswj/2014/373254.pdf · Ternary code ( 1 )1101 ( 1 )0( 1 ) F : LTP operator. where , , ,and

Research ArticleCompleted Local Ternary Pattern for RotationInvariant Texture Classification

Taha H Rassem and Bee Ee Khoo

School of Electrical amp Electronic Engineering Universiti Sains Malaysia Engineering Campus Nibong Tebal 14300 Penang Malaysia

Correspondence should be addressed to Bee Ee Khoo beekhoousmmy

Received 20 December 2013 Accepted 11 February 2014 Published 7 April 2014

Academic Editors G C Gini and L Li

Copyright copy 2014 T H Rassem and B E KhooThis is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in anymedium provided the originalwork is properly cited

Despite the fact that the two texture descriptors the completed modeling of Local Binary Pattern (CLBP) and the CompletedLocal Binary Count (CLBC) have achieved a remarkable accuracy for invariant rotation texture classification they inherit someLocal Binary Pattern (LBP) drawbacks The LBP is sensitive to noise and different patterns of LBP may be classified into thesame class that reduces its discriminating property Although the Local Ternary Pattern (LTP) is proposed to be more robustto noise than LBP however the latterrsquos weakness may appear with the LTP as well as with LBP In this paper a novel completedmodeling of the Local Ternary Pattern (LTP) operator is proposed to overcome both LBP drawbacks and an associated completedLocal Ternary Pattern (CLTP) scheme is developed for rotation invariant texture classificationThe experimental results using fourdifferent texture databases show that the proposed CLTP achieved an impressive classification accuracy as compared to the CLBPand CLBC descriptors

1 Introduction

Nowadays texture analysis and classification have becomeone of the important areas of computer vision and imageprocessingThey play a vital role inmany applications such asvisual object recognition and detection [1 2] human detector[3] object tracking [4] pedestrian classification [5] imageretrieval [6 7] and face recognition [8 9]

Currently many textures feature extraction algorithmsthat have been proposed to achieve a good texture classifica-tion Most of these algorithms are focusing on how to extractdistinctive texture features that are robust to noise rotationand illumination variance These algorithms can be classifiedinto three categories [10] The first category is the statisticalmethods such as polar plots and polarograms [11] textureedge statistics on polar plots [12]moment invariants [13] andfeature distribution method [14] The second category is themodel based methods such as simultaneous autoregressivemodel (SAR) [15] Markov model [16] and steerable pyramid[17] The third category is the structural methods suchas topological texture descriptors [18] invariant histogram[19] and morphological decomposition [20] All of thesealgorithms as well as many other algorithms are reviewedbriefly in many review papers [10 21 22]

Local Binary Pattern (LBP) operator was proposed byOjala et al [23] for rotation invariant texture classificationIt has been modified and adapted for several applicationssuch as face recognition [8 9] and image retrieval [7] TheLBP extraction algorithm contains two main steps that isthe thresholding step and the encoding step This is shownin Figure 1 In the thresholding step all the neighboringpixel values in each pattern are compared with the value oftheir central pixel of the pattern to convert their values tobinary values (0 or 1) This step helps to get the informationabout the local binary differences Then in the encodingstep the binary numbers obtained from the thresholdingstep are encoded and converted into a decimal number tocharacterize a structural pattern In the beginning Ojala etal [24] represented the texture image using textons histogramby calculating the absolute difference between the gray levelof the center pixel of a specific local pattern and its neighborsThen the authors proposed the LBP operator by using the signof the differences between the gray level of the center pixeland its neighbors of the local pattern instead of magnitude[23] LBP proposed by Ojala et al [23] has become theresearch direction formany computer vision researchersThisis because it is able to distinguish the microstructures suchas edges lines and spots The researchers aim to increasethe discriminating property of the texture feature extraction

Hindawi Publishing Corporatione Scientific World JournalVolume 2014 Article ID 373254 10 pageshttpdxdoiorg1011552014373254

2 The Scientific World Journal

52 90 102

70 34 20

12 40 9

1 1 1

1 0

0 1 0

34

Threshold Encoding

Binary01111010

Decimal122

Figure 1 LBP operator

to achieve impressive rotation invariant texture classificationSo many of the variants of the LBP have been suggestedand proposed for rotation invariant texture classificationThecenter-symmetric Local Binary Pattern (CS-LBP) proposedby Heikkil et al [25] is an example for that Unlike the LBPthey compared center-symmetric pairs of pixels to get theencoded binary values Liao et al [26] proposed DominantLBP (DLBP) by selecting the dominant patterns from allrotation invariant patterns Tan and Triggs [27] presenteda new texture operator which is more robust to noiseThey encoded the neighbor pixel values into 3-valued codes(minus1 0 1) instead of 2-valued codes (0 1) by adding a userthreshold This operator is known as a Local Ternary Pattern(LTP) Guo et al [28] combined the sign and magnitudedifferences of each pattern with all central gray level valuesof all patterns to propose a completed modeling of LBPcalled completed LBP (CLBP) Khellah [29] proposed a newmethod for texture classification which combines DominantNeighborhood Structure (DNS) and traditional LBP Zhaoet al [30] proposed a novel texture descriptor called LocalBinary Count (CLBC) They used the thresholding step suchas in LBP Then they discarded the structural informationfrom the LBP operator by counting the number of value 1rsquosin the binary neighbor sets instead of encoding them

Although some LBP variant descriptors such as CLBPand CLBC have achieved remarkable classification accuracythey inherit the LBP weaknesses The LBP suffers from twomain weaknesses It is sensitive to noise and sometimes mayclassify two or more different patterns falsely to the sameclass as shown in Figures 2 and 3The LTP descriptor is morerobust to noise than LBP However the latter weakness mayappear with the LTP as well as with LBP

In this paper we are enhancing the LTP texture descrip-tor to increase its discriminating property by presenting acompleted modeling for LTP operator and proposing anassociated completed Local Ternary Pattern (CLTP) Theexperimental results illustrate that CLTP is more robust tonoise rotation and illumination variance and it achieveshigher texture classification rates than CLBP and CLBC Therest of this paper is organized as follows Section 2 brieflyreviews the LBP LTP CLBP and CLBC Section 3 presentsthe new CLTP scheme Then in Section 4 the experimentalresults of different texture databases are reported and dis-cussed Finally Section 5 concludes the paper

2 Related Work

In this section a brief review of the LBP LTP CLBP andCLBC is provided

46 46 45

51 52 52

50 52 53

46 46 45

51 52

50 52 53

56

LBP 00000000LBP 100000011

Noise

52 to 56

Figure 2 The example for LBP operatorrsquos noise sensitivity

52 90 102

70 34 20

12 40 9

130 200 95

92 90 70

12 100 9

LBP 01111010 LBP 01111010

Figure 3 Similar LBP codes for two different texture patterns

21 Brief Review of LBP As shown in Figure 1 the LBPoperator is computed for the center pixel by comparing theintensity value of it with the intensity values of its neighborsThis process can be expressed mathematically as follows

LBP119875119877

=

119875minus1

sum119901=0

2119901119904 (119894119901minus 119894119888) 119904 (119909) =

1 119909 ge 0

0 119909 lt 0(1)

where 119894119888and 119894119901(119901 = 0 119875 minus 1) denote the gray value of the

center pixel and gray value of the neighbor pixel on a circle ofradius 119877 respectively and 119875 is the number of the neighborsBilinear interpolation estimation method is used to estimatethe neighbors that do not lie exactly in the center of the pixelsLBPri119875119877

and LBPriu2119875119877

are rotation invariant of LBP and uniformrotation invariant of LBP respectively These two enhancedLBP operators are proposed by Ojala et al [23]

After completing the encoding step for any LBP opera-tors that is LBPri

119875119877and LBPriu2

119875119877 the histogram can be created

based on the following equation

119867(119896) =

119868

sum119894=0

119869

sum119895=0

119891 (LBP119875119877

(119894 119895) 119896) 119896 isin [0 119870]

119891 (119909 119910) = 1 119909 = 119910

0 otherwise

(2)

where119870 is the maximal LBP pattern value

22 Brief Review of LTP Tan and Triggs [27] presented anew texture operator which is more robust to noise The LBPis extended to 3-valued codes (minus1 0 1) Figure 4 shows anexample of LTP operatorThemathematical expression of theLTP can be described as follows

LTP119875119877

=

119875minus1

sum119901=0

2119901119904 (119894119901minus 119894119888) 119904 (119909) =

1 119909 ge 119905

0 minus119905 lt 119909 lt 119905

minus1 119909 lt minus119905

(3)

The Scientific World Journal 3

31 90 102

70 34 20

12 33 9

0

0

1 1

1

0

00

0

0

0

00 0

0

1

1

1

1

1

1

minus1

minus1minus1

Upper pattern

Lower pattern

t = 5

Threshold

[34 minus t 34 + t]

Binary01101000

Decimal104

Binary10000101

Decimal133

Ternary code(minus1)1101(minus1)0(minus1)

Figure 4 LTP operator

where 119894119888 119894119901 119877 and 119875 are defined before in (1) and 119905 denotes

the user threshold After thresholding step the upper patternand lower pattern are constructed and coded as shown inFigure 4 The LTP operator is the concatenation of the codeof the upper pattern and the lower pattern

23 Brief Review of Completed LBP (CLBP) The completedLBP (CLBP) descriptor was proposed by Guo et al [28]in order to improve the performance of LBP descriptor Asshown in Figure 5 the image local difference is decomposedinto two complementary components the sign component 119904

119901

and the magnitude component119898119901 Consider the following

119904119901= 119904 (119894119901minus 119894119888) 119898

119901=10038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816 (4)

Then the 119904119901is used to build the CLBP-Sign (CLBP S)

whereas the119898119901is used to build CLBP-magnitude (CLBP M)

The CLBP S and CLBP M are described mathematically asfollows

CLBP S119875119877

=

119875minus1

sum119901=0

2119901119904 (119894119901minus 119894119888) 119904

119901=

1 119894119901ge 119894119888

0 119894119901lt 119894119888

(5)

CLBP M119875119877

=

119875minus1

sum119901=0

2119901119905 (119898119901 119888)

119905 (119898119901 119888) =

110038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816lt 119888

(6)

where 119894119888 119894119901 119877 and 119875 are defined before in (1) and 119888 in (6)

denotes the mean value of119898119901in the whole image

The CLBP S is equal to LBP whereas the CLBP Mmeasures the local variance of magnitude Furthermore Guoet al used the value of the gray level of each pattern toconstruct a newoperator calledCLBP-Center (CLBP C)TheCLBP C can be described mathematically as follows

CLBP C119875119877

= 119905 (119894119888 119888119868) (7)

where 119894119888denotes the gray value of the center pixel and the 119888

119868is

the average gray level of thewhole image Guo et al combinedtheir operators into joint or hybrid distributions and achieveda remarkable texture classification accuracy [28]

24 Brief Review of Completed LBC (CLBC) The LocalBinary Count (LBC) was proposed by Zhao et al [30]Unlike the LBP and all its variants the authors just countedthe number of value 1rsquos of the thresholding step instead ofencoding themThe LBC can be described mathematically asfollows

LBC119875119877

=

119875minus1

sum119901=0

119904 (119894119901minus 119894119888) 119904 (119909) =

1 119909 ge 0

0 119909 lt 0(8)

Similar to CLBP the authors [30] extended the LBC to com-pleted LBC (CLBC) The CLBC S CLBC M and CLBC Cwere also combined into joint or hybrid distributions andthey were used for rotation invariant texture classificationThe CLBC M and CLBC C can be described mathematicallyas follows

CLBC M119875119877

=

119875minus1

sum119901=0

119905 (119898119901 119888) 119905 (119898

119901 119888) =

110038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816lt 119888

CLBC C119875119877

= 119905 (119894119888 119888119868)

(9)

where 119894119888 119894119901 119888 and 119888

119868are defined in (1) (6) and (7) An

example of LBC is shown in Figure 6In [28] the rotation invariant LBP (LBPriu2

119875119877) is used to

construct the CLBPriu2119875119877

The CLBPriu2119875119877

is simplified in thispaper as CLBP

119875119877as well as the proposed CLTP operator

3 Completed Local Ternary Pattern (CLTP)

In this section we propose the framework of CLTP Similar toCLBP [28] the LTP is extended to completed modeling LTP


31 90 102

70 34 20

12 33 9

0 1 1

1 0

000

LBP S = 01101000

(a)

3 56 68

36 14

22 1 25

0 1 1

1 0

000

Assume c = 29

LBP M = 01101000

(b)

Figure 5 A 3times3 sample pattern (a) sign component (LBP S code)(b) magnitude components (LBP M code (assume threshold = 29))

52 90 102

70 34 20

12 40 9

1 1 1

1 0

0 1 0

34

ThresholdLBC code 5

Figure 6 LBC operator

(CLTP) Asmentioned before the LTP ismore robust to noisethan LBP Furthermore construct the associated completedLocal Ternary Pattern that will help to enhance and increaseits the discriminating property The mathematical model ofLTP is shown in (3) In CLTP local difference of the image isdecomposed into two sign complementary components andtwo magnitude complementary components as follows

119904upper119901

= 119904 (119894119901minus (119894119888+ 119905)) 119904

lower119901

= 119904 (119894119901minus (119894119888minus 119905))

119898upper119901

=10038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816 119898

lower119901

=10038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816

(10)

where 119894119888 119894119901 and 119905 are defined before in (1) and (3)

Then the 119904upper119901 and 119904lower119901

are used to build theCLTP Supper119875119877

and CLTP Slower119875119877

respectively as follows

CLTP Supper119875119877

=

119875minus1

sum119901=0

2119901119904 (119894119901minus (119894119888+ 119905))

119904upper119901

= 1 119894119901ge 119894119888+ 119905

0 otherwise

CLTP Slower119875119877

=

119875minus1

sum119901=0

2119901119904 (119894119901minus (119894119888minus 119905))

119904lower119901

= 1 119894119901lt 119894119888minus 119905

0 otherwise(11)

The CLTP S119875119877

is the concatenation of the CLTP Supper119875119877


as follows

CLTP S119875119877

= [CLTP Supper119875119877


] (12)

where 119894119888 119894119901 119875 119877 and 119905 in (11) are defined before in (3)

Similar to CLTP S119875119877

the CLTP M119875119877

is built usingthe two magnitude complementary components 119898upper

119901 and119898lower119901

as follows

CLTP Mupper119875119877

=

119875minus1

sum119901=0

2119901119905 (119898

upper119901

119888)

119905 (119898upper119901

119888) =

110038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816lt 119888

(13)

CLTP Mlower119875119877

=

119875minus1

sum119901=0

2119901119905 (119898

lower119901

119888)

119905 (119898lower119901

119888) =

110038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816lt 119888

(14)

CLTP M119875119877

= [CLTP Mupper119875119877


] (15)

where 119894119888 119894119901 119875 119877 and 119905 in (13) and (14) are defined before in

(3) and 119888 is defined in (6)Moreover the CLTP Cupper

119875119877and CLTP Clower

119875119877can be

mathematically described as follows

CLTP Cupper119875119877

= 119905 (119894upper119888

119888119868)

CLTP Clower119875119877

= 119905 (119894lower119888

119888119868)

(16)

where 119894upper119888

= 119894119888+ 119905 119894lower119888

= 119894119888minus 119905 and 119888

119868is the average gray

level of the whole imageThe proposed CLTP operators are combined into joint

or hybrid distributions to build the final operator histogramlike the CLBP and CLBC [28 30] respectively In the CLTPthe operators of the same type of pattern that is the upperand the lower pattern are combined first into joint or hybriddistributionsThen their results are concatenated to build thefinal operator histogramThat mean number of bins of CLTPis double in size than the number of bins of CLBP

4 Experiments and Discussion

In this section a series of experiments are performed toevaluate the proposed CLTP Four large and comprehensivetexture databases are used in these experiments They arethe Outex database [31] Columbia-Utrecht Reflection andTexture (CUReT) database [32] UIUC database [33] andXU HR database [34] Empirically the threshold value 119905 isset to 5


41 Dissimilarity Measuring Framework Several metrics areproposed for measuring the dissimilarity between two his-tograms such as log-likelihood ratio histogram intersectionand chi-square statistic Similar to [28 30] the chi-squarestatistic is used in this brief The 1205942 distance between twohistograms119867 = ℎ

119894and119870 = 119896

119894where (119894 = 1 2 3 119861) can be


Dissimilarity1205942 (119867119870) =

119861

sum119894=1

(ℎ119894minus 119896119894)2

ℎ119894+ 119896119894

(17)

Furthermore the nearest neighborhood classifier is used forclassification in all experiments in this paper

42 Experimental Results on the Outex Database The Outexdatasets (httpwwwoutexoulufiindexphppage=classifi-cation) include 16 test suites starting from Outex TC 00010(TC10) to Outex TC 00016 (TC16) [31] These suites werecollected under different illumination rotation and scalingconditions Outex TC 00010 (TC10) and Outex TC 00012(TC12) are considered as famous two test suites in Outexdatasets These two suites have the same 24 classes of tex-tures which were collected under three different illuminates(ldquohorizonrdquo ldquoincardquo and ldquot184rdquo) and nine different rotationangles (0∘ 5∘ 10∘ 15∘ 30∘ 45∘ 60∘ 75∘ and 90∘) Foreach illumination and rotation situation each class has 20nonoverlapping texture samples with size of 128 times 128Examples of Outex images are shown in Figure 7 For TC10480 images are used as training data These are the imagesof ldquoincardquo illumination condition and ldquo0∘rdquo angle rotationwhereas the images under the remaining rotation angles andldquoincardquo illumination condition are used as testing data that is3840 images The training data in case of TC12 is the same asTC10 while all images under ldquot184rdquo or ldquohorizonrdquo illuminationconditions are used as testing data that is 4320 imagesfor ldquot184rdquo and 4320 images for ldquohorizonrdquo The experimentalresults of TC10 TC12 (t184) and TC12 (horizon) are shownin Table 1

From Table 1 the following points can be observedFirstly the CLTP S CLTP M CLTP MC and CLTP S MCperformed better than the similar CLBP and CLBC typesoperators Secondly the CLTP S and CLTP M showed reallygreat discrimination capability than CLBP S and CLBP Mand CLBC S and CLBC M respectively where the accu-racy difference exceeded 10 in some cases Thirdly theCLTP SM worked well with TC10 and TC12 for (119875 = 1 and119877 = 8) with TC12 for (119875 = 2 and 119877 = 16) and only withTC12 under ldquot184rdquo illumination condition for (119875 = 3 and 119877 =

24) Finally the CLTP SMC achieved the best classificationaccuracy with TC10 and TC12 for (119875 = 2 and 119877 = 16)and TC10 and TC12 under ldquot184rdquo illumination condition for(119875 = 3 and 119877 = 24)

43 Experimental Results on CUReT Database The CUReTdataset (httpwwwrobotsoxacuksimvggresearchtexclassindexhtml) has 61 texture classes In each class there are205 images subjected to different illumination and viewpointsconditions [32] The images in each class have differentviewing angle shots Out of 250 images in each class there

are 118 image shots whose viewing angles are lesser than60∘ Examples of CUReT images are shown in Figure 8

From these types of images only 92 images are selectedafter converting to gray scale and cropping to 200 times 200In each class 119873 images from 92 are used as training datawhile the remaining (92 minus 119873) are used as testing data Thefinal classification accuracy is the average percentage over ahundred random splitsThe CUReT average classification for119873 = (6 12 23 46) is shown in Table 2 It is easier to note thatCLTP has better performance than CLBP and CLBC with theCUReT database Except CLTP SMC all CLTP operatorsare achieving higher classification rates than other CLBP andCLBC for all cases of 119873 images and at every radius TheCLTP SMC achieved the best classification rates for all 119873training images at radius 3 for 6 23 and 64 training imagesat radius 2 and for 6 training images only at radius 1 whileat radius 1 the CLBP SMC achieved the best classificationrates for 12 23 and 64 training images and for 12 trainingimages at radius 2

44 Experimental Results on UIUC Database The UIUCdatabase has 25 classes containing the images captured undersignificant viewpoint variations In each class there are 40images with resolution of 640 times 480 Examples of UIUCimages are shown in Figure 9 In each class 119873 images from40 are used as training data while the remaining (40 minus 119873)

are used as testing data The final classification accuracyis the average percentage over a hundred random splitsThe UIUC average classification for 119873 = (5 10 15 20) isshown in Table 3 Except in case of CLTP SMC all CLTPoperators achieved a higher performance than CLBP andCLBC operators for all 119873 number of training images atradiuses 1 2 and 3 The CLTP SMC outperformed theCLBP SMC and CLBC SMC for all119873 number of trainingimages when 119875 = 3 119877 = 24 and 119875 = 2 119877 = 16 On the otherhand the CLBC SMC is the best one for small number of119873 that is 5 and 10 CLBP SMC is the best one for119873 = 20

and the CLTP SMC is the best one for119873 = 15

45 Experimental Results on XU-HR Database The XU HRdatabase has 25 classes with 40 high resolution images (1280times960) in each class Examples of XU-HR images are shown inFigure 10 In each class119873 images from40 are used as trainingdata while the remaining (40 minus 119873) are used as testing dataThefinal classification accuracy is the average percentage overa hundred random splits The XU HR average classificationfor 119873 = (5 10 15 20) is shown in Table 4 In XU HRdatabase the CLTP achieved higher classification rates thanCLBP for all 119873 training images with all types of radiusesThe CLTP SMC achieved an impressive classification ratereaching 99 when119873 = 20 at (119875 = 3 119877 = 24) We comparedthis database only with CLBP since HU HR results usingCLBC are not available in [30]

5 Conclusion

To overcome some drawbacks of LBP an existing LTPoperator is extended to build a new texture operator definedas completed Local Ternary PatternThe proposed associated


Table 1 Classification rates () on TC10 and TC12 database

119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8

TC10 TC12 Average TC10 TC12 Average TC10 TC12 Averaget184 Horizon t184 Horizon t184 Horizon

CLBP S 9507 8504 8078 8696 8940 8226 7520 8229 8441 6546 6368 7118CLBC S [30] 9135 8382 8275 8597 8867 8257 7741 8288 8294 6502 6317 7038CLTP S 9820 9359 8942 9374 9695 9016 8694 9135 9414 7588 7396 8133CLBP M 9552 8118 7865 8512 9367 7379 7240 7995 8174 5930 6277 6794CLBC M [30] 9185 7559 7458 8067 9245 7035 7264 7848 7896 5363 5801 6353CLTP M 9800 8539 8465 8935 9732 8340 8440 8837 9404 7586 7405 8132CLBP MC 9802 9074 9069 9315 9744 8694 9097 9178 9036 7238 7666 7980CLTP MC 9852 9123 8998 9324 9794 9014 9238 9349 9594 8470 8602 8889CLBP S MC 9833 9405 9240 9493 9802 9099 9108 9336 9453 8187 8252 8631CLTP S MC 9898 9500 9294 9564 9844 9241 9280 9455 9643 8400 8685 8909CLBP SM 9932 9358 9335 9542 9789 9055 9111 9318 9466 8275 8314 8685CLBC SM [30] 9870 9141 9025 9345 9810 8995 9042 9282 9523 8213 8359 8698CLTP SM 9904 9414 9259 9526 9784 9206 9269 9420 9641 8285 8481 8802CLBP SMC 9893 9532 9453 9626 9872 9354 9391 9539 9656 9030 9229 9305CLBC SMC [30] 9878 9400 9324 9534 9854 9326 9407 9529 9716 8979 9292 9329CLBC CLBP [30] 9896 9537 9472 9635 9883 9359 9426 9556 9688 9025 9292 9335DLBPP [30] 9810 9160 8740 9237 9770 9210 8870 9283 mdash mdash mdash mdashCLTP SMC 9917 9567 9428 9637 9893 9403 9479 9592 9698 8706 9030 9145The bold values indicate higher classification rate

Figure 7 Some images from Outex database


Table 2 Classification rates () on CUReT database

119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8

6 12 23 46 6 12 23 46 6 12 23 46CLBP S [28] 6694 7526 8180 8731 6349 7268 7949 8535 5900 6781 7462 8070CLBC S [30] 6082 7057 7421 8014 6042 6895 7442 7978 5688 6621 7289 7882CLTP S 7257 8155 8772 9175 6839 7909 8661 9155 6438 7266 8173 8824CLBP M [28] 6257 7193 7989 8649 5857 6828 7611 8303 5177 6033 6773 7516CLBC M [30] 5123 6053 6836 7741 5063 5870 6605 7389 5012 5862 5782 6661CLTP M 6714 7693 8516 9052 6333 7447 8214 8883 6137 7117 8053 8667CLBP MC [28] 6871 7854 8604 9165 6481 7556 8298 8975 5653 6715 7558 8297CLTP MC 7010 8012 8902 9358 6677 7712 8551 9167 6207 7294 8226 8898CLBP S MC [28] 7329 8228 8928 9407 7027 8047 8757 9278 6663 7654 8502 9055CLTP S MC 7436 8514 9103 9469 7155 8216 8782 9404 6754 7889 8546 9127CLBP SM [28] 7495 8430 9083 9453 7463 8344 8967 9385 7186 8227 8857 9346CLBC SM [30] 7052 8157 8912 9360 7216 8271 8960 9378 6989 7988 8662 9310CLTP SM 7649 8511 9202 9563 7414 8442 9078 9506 7130 8237 8920 9350CLBP SMC [28] 7680 8654 9200 9572 7607 8573 9215 9567 7435 8506 9152 9570CLBC SMC [30] 7318 8407 9055 9526 7517 8591 9130 9539 7285 8292 9012 9478CLTP SMC 7797 8750 9272 9611 7772 8554 9244 9595 7518 8406 9045 9478The bold values indicate higher classification rate

Table 3 Classification rates () on UIUC database

119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8

5 10 15 20 5 10 15 20 5 10 15 20CLBP S 4487 5468 6063 6420 4180 5134 5680 6060 4005 4753 5163 5529CLBC S [30] 4719 5746 6348 6690 4337 5307 5917 6239 3985 4669 5111 5561CLTP S 6880 7760 8304 8600 6491 7507 8048 8320 5429 6187 6992 7160CLBP M 5615 6592 7105 7437 5607 6565 6951 7205 4239 4998 5445 5752CLBC M [30] 5168 6062 6663 6933 5067 5901 6442 6710 3904 4551 4942 5212CLTP M 6994 7933 8256 8520 7029 7933 8336 8540 5749 6467 6960 7360CLBP MC 6808 7675 8081 8327 6845 7683 8014 8272 5692 6509 6981 7266CLTP MC 7680 8347 8720 8860 7737 8360 8704 8940 7006 7693 8048 8180CLBP S MC 6943 7861 8281 8533 6868 7757 8136 8355 6252 7127 7548 7865CLTP S MC 7726 8467 8848 9060 7737 8427 8784 8980 6880 7733 8048 8360CLBP SM 7205 8263 8688 8956 7180 8085 8531 8760 6470 7465 7955 8258CLBC SM [30] 7516 8392 8768 8972 7316 8204 8631 8851 6528 7488 7886 8240CLTP SM 7931 8773 9056 9320 7714 8560 8944 9180 6503 7440 7968 8300CLBP SMC 7805 8587 8917 9107 7875 8633 8925 9103 7453 8226 8585 8786CLBC SMC [30] 7975 8645 9010 9139 7948 8663 8966 9104 7457 8235 8566 8783CLTP SMC 8297 8893 9152 9440 8263 8787 9040 9260 7451 8173 8592 8680The bold values indicate higher classification rate

Table 4 Classification rates () on XU HR database

119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8

5 10 15 20 5 10 15 20 5 10 15 20CLBP S 7552 8427 8862 9114 7411 8295 8644 8877 7353 8197 8591 8847CLTP S 8606 9333 9520 9660 8446 9067 9360 9420 7680 8413 8928 9120CLBP M 7769 8527 8907 9146 7775 8499 8854 9058 6813 7629 8058 8364CLTP M 8651 9213 9472 9620 8423 9093 9344 9480 7691 8413 8832 9120CLBP MC 8589 9129 9369 9520 8580 9132 9364 9507 8121 8757 9040 9211CLTP MC 9131 9627 9760 9760 9029 9547 9696 9800 8800 9240 9472 9600CLBP S MC 8726 9254 9467 9606 8723 9266 9472 9598 8637 9160 9357 9465CLTP S MC 9189 9667 9712 9820 9063 9507 9680 9860 8766 9320 9552 9620CLBP SM 8709 9266 9482 9610 8698 9245 9467 9609 8427 9024 9230 9376CLTP SM 9177 9560 9792 9860 9051 9533 9696 9800 8423 9040 9328 9480CLBP SMC 9111 9533 9666 9747 9164 9521 9680 9750 9081 9477 9580 9683CLTP SMC 9554 9800 9856 9900 9417 9733 9856 9880 9120 9547 9648 9780The bold values indicate higher classification rate


Figure 8 Some images from CUReT dataset

Figure 9 Some images from UIUC database


Figure 10 Some images from XU-HR database

completed Local Ternary Pattern (CLTP) scheme is evalu-ated using four challenging texture databases for rotationinvariant texture classification The experimental results inthis paper demonstrate the superiority of the proposed CLTPagainst the new existing texture operators that is CLBP andCLBC This is because the proposed CLTP is insensitive tonoise and has a high discriminating property that leads toachieve impressive classification accuracy rates

Conflict of Interests

The authors declare that there is no conflict of interests regar-ding the publication of this paper

Acknowledgment

The authors would like to sincerely thank Z Guo for sharingthe source codes of CLBP

References

[1] J Xiao J Hays K A Ehinger A Oliva and A Torralba ldquoSUNdatabase large-scale scene recognition from abbey to zoordquo inProceedings of the IEEE Computer Society Conference on Com-puter Vision and Pattern Recognition (CVPR rsquo10) pp 3485ndash3492June 2010

[2] Y Lee D K Han and H Ko ldquoReinforced adaboost learning forobject detection with local pattern representationsrdquo The Scie-ntific World Journal vol 2013 Article ID 153465 14 pages 2013

[3] X Wang T Han and S Yan ldquoAn HOG-LBP human detectorwith partial occlusion handlingrdquo in Proceedings of the 12th IEEEInternational Conference of Computer Vision pp 32ndash39 2009

[4] V Takala and M Pietikainen ldquoMulti-object tracking usingcolor texture and motionrdquo in Proceedings of the IEEE ComputerSociety Conference on Computer Vision and Pattern Recognition(CVPR rsquo07) pp 1ndash7 June 2007

[5] M Enzweiler and D M Gavrila ldquoA multilevel mixture-of-exp-erts framework for pedestrian classificationrdquo IEEE Transactionson Image Processing vol 20 no 10 pp 2967ndash2979 2011

[6] F Qiao C Wang X Zhang and H Wang ldquoLarge scale near-duplicate celebrity web images retrieval using visual and textualfeaturesrdquo The Scientific World Journal vol 2013 Article ID795408 11 pages 2013

[7] S Murala R P Maheshwari and R Balasubramanian ldquoLocaltetra patterns a new feature descriptor for content-based imageretrievalrdquo IEEE Transactions on Image Processing vol 21 no 5pp 2874ndash2886 2012

[8] B Zhang Y Gao S Zhao and J Liu ldquoLocal derivative patternversus local binary pattern face recognition with high-orderlocal pattern descriptorrdquo IEEETransactions on Image Processingvol 19 no 2 pp 533ndash544 2010

[9] J Y Choi Y M Ro and K N Plataniotis ldquoColor local texturefeatures for color face recognitionrdquo IEEE Transactions on ImageProcessing vol 21 no 3 pp 1366ndash1380 2012


[10] J Zhang and T Tan ldquoBrief review of invariant texture analysismethodsrdquo Pattern Recognition vol 35 no 3 pp 735ndash747 2002

[11] L S Davis ldquoPolarograms a new tool for image texture analysisrdquoPattern Recognition vol 13 no 3 pp 219ndash223 1981

[12] M AMayorga and L C Ludeman ldquoShift and rotation invarianttexture recognition with neural netsrdquo in Proceedings of theIEEE International Conference on Neural Networks IEEEWorldCongress on Computational Intelligence vol 6 pp 4078ndash4083June 1994

[13] M K Hu ldquoVisual pattern recognition by moment invariantsrdquoIRE Transactions on Information Theory vol 8 no 2 pp 179ndash187 1962

[14] M Pietikainen T Ojala and Z Xu ldquoRotation-invariant textureclassification using feature distributionsrdquo Pattern Recognitionvol 33 no 1 pp 43ndash52 2000

[15] R L Kashyap and A Khotanzad ldquoA model-based method forrotation invariant texture classificationrdquo IEEE Transactions onPattern Analysis and Machine Intelligence vol 8 no 4 pp 472ndash481 1986

[16] F S Cohen Z Fan and M A Patel ldquoClassification of rotatedand scaled textured images using Gaussian Markov randomfield modelsrdquo IEEE Transactions on Pattern Analysis and Mac-hine Intelligence vol 13 no 2 pp 192ndash202 1991

[17] H Greenspan S Belongie R Goodman and P Perona ldquoRota-tion invariant texture recognition using a steerable pyramidrdquo inProceedings of the 12th International Conference on PatternRecognition vol 2 pp 162ndash167 1994

[18] G Eichmann and T Kasparis ldquoTopologically invariant texturedescriptorsrdquo Computer Vision Graphics and Image Processingvol 41 no 3 pp 267ndash281 1988

[19] R K Goyal W L Goh D P Mital and K L Chan ldquoScale androtation invariant texture analysis based on structural propertyrdquoin Proceedings of the IEEE 21st International Conference on Ind-ustrial Electronics Control and Instrumentation vol 2 pp1290ndash1294 November 1995

[20] W K Lam and C Li ldquoRotated texture classification by imp-roved iterative morphological decompositionrdquo IEE ProceedingsVision Image and Signal Processing vol 144 no 3 pp 171ndash1791997

[21] T R Reed and J M H Dubuf ldquoA review of recent texture seg-mentation and feature extraction techniquesrdquo CVGIP ImageUnderstanding vol 57 no 3 pp 359ndash372 1993

[22] R W Conners and C A Harlow ldquoA theoretical comparison oftexture algorithmsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 2 no 3 pp 204ndash222 1980

[23] 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

[24] TOjalaM Pietikainen andDHarwood ldquoA comparative studyof texture measures with classification based on feature distri-butionsrdquo Pattern Recognition vol 29 no 1 pp 51ndash59 1996

[25] MHeikkilM Pietikinen andC Schmid ldquoDescription of inter-est regions with center-symmetric local binary patternsrdquo in Pro-ceedings of 5th Indian Conference of Computer Vision Graphicsand Image Processing vol 4338 pp 58ndash69 2006

[26] S LiaoMW K Law and A C S Chung ldquoDominant local bin-ary patterns for texture classificationrdquo IEEE Transactions onImage Processing vol 18 no 5 pp 1107ndash1118 2009

[27] X Tan and B Triggs ldquoEnhanced local texture feature sets forface recognition under difficult lighting conditionsrdquo IEEE Tran-sactions on Image Processing vol 19 no 6 pp 1635ndash1650 2010

[28] ZGuo L Zhang andD Zhang ldquoA completedmodeling of localbinary pattern operator for texture classificationrdquo IEEE Trans-actions on Image Processing vol 19 no 6 pp 1657ndash1663 2010

[29] FM Khellah ldquoTexture classification using dominant neighbor-hood structurerdquo IEEE Transactions on Image Processing vol 20no 11 pp 3270ndash3279 2011

[30] Y Zhao D S Huang andW Jia ldquoCompleted local binary countfor rotation invariant texture classificationrdquo IEEE Transactionson Image Processing vol 21 no 10 pp 4492ndash4497 2012

[31] TOjala TMaenpaaM Pietikainen J Viertola J Kyllonen andS Huovinen ldquoOutexmdashnew framework for empirical evaluationof texture analysis algorithmsrdquo in Proceedings of the 16th Inter-national Conference on Pattern Recognition no 1 pp 701ndash77062002

[32] M Varma and A Zisserman ldquoClassifying images of materialsachieving viewpoint and illumination independencerdquo in Pro-ceedings of 7th European Conference on Computer Vision vol3 pp 255ndash271 May 2002

[33] S Lazebnik C Schmid and J Ponce ldquoA sparse texture represen-tation using local affine regionsrdquo IEEE Transactions on PatternAnalysis and Machine Intelligence vol 27 no 8 pp 1265ndash12782005

[34] Y Xu H Ji and C Fermuller ldquoViewpoint invariant texture des-cription using fractal analysisrdquo International Journal of Com-puter Vision vol 83 no 1 pp 85ndash100 2009

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



52 90 102

70 34 20

12 40 9

1 1 1

1 0

0 1 0

34

Threshold Encoding

Binary01111010

Decimal122

Figure 1 LBP operator

to achieve impressive rotation invariant texture classificationSo many of the variants of the LBP have been suggestedand proposed for rotation invariant texture classificationThecenter-symmetric Local Binary Pattern (CS-LBP) proposedby Heikkil et al [25] is an example for that Unlike the LBPthey compared center-symmetric pairs of pixels to get theencoded binary values Liao et al [26] proposed DominantLBP (DLBP) by selecting the dominant patterns from allrotation invariant patterns Tan and Triggs [27] presenteda new texture operator which is more robust to noiseThey encoded the neighbor pixel values into 3-valued codes(minus1 0 1) instead of 2-valued codes (0 1) by adding a userthreshold This operator is known as a Local Ternary Pattern(LTP) Guo et al [28] combined the sign and magnitudedifferences of each pattern with all central gray level valuesof all patterns to propose a completed modeling of LBPcalled completed LBP (CLBP) Khellah [29] proposed a newmethod for texture classification which combines DominantNeighborhood Structure (DNS) and traditional LBP Zhaoet al [30] proposed a novel texture descriptor called LocalBinary Count (CLBC) They used the thresholding step suchas in LBP Then they discarded the structural informationfrom the LBP operator by counting the number of value 1rsquosin the binary neighbor sets instead of encoding them

Although some LBP variant descriptors such as CLBPand CLBC have achieved remarkable classification accuracythey inherit the LBP weaknesses The LBP suffers from twomain weaknesses It is sensitive to noise and sometimes mayclassify two or more different patterns falsely to the sameclass as shown in Figures 2 and 3The LTP descriptor is morerobust to noise than LBP However the latter weakness mayappear with the LTP as well as with LBP

In this paper we are enhancing the LTP texture descrip-tor to increase its discriminating property by presenting acompleted modeling for LTP operator and proposing anassociated completed Local Ternary Pattern (CLTP) Theexperimental results illustrate that CLTP is more robust tonoise rotation and illumination variance and it achieveshigher texture classification rates than CLBP and CLBC Therest of this paper is organized as follows Section 2 brieflyreviews the LBP LTP CLBP and CLBC Section 3 presentsthe new CLTP scheme Then in Section 4 the experimentalresults of different texture databases are reported and dis-cussed Finally Section 5 concludes the paper

2 Related Work

In this section a brief review of the LBP LTP CLBP andCLBC is provided

46 46 45

51 52 52

50 52 53

46 46 45

51 52

50 52 53

56

LBP 00000000LBP 100000011

Noise

52 to 56

Figure 2 The example for LBP operatorrsquos noise sensitivity

52 90 102

70 34 20

12 40 9

130 200 95

92 90 70

12 100 9

LBP 01111010 LBP 01111010

Figure 3 Similar LBP codes for two different texture patterns

21 Brief Review of LBP As shown in Figure 1 the LBPoperator is computed for the center pixel by comparing theintensity value of it with the intensity values of its neighborsThis process can be expressed mathematically as follows

LBP119875119877

=

119875minus1

sum119901=0

2119901119904 (119894119901minus 119894119888) 119904 (119909) =

1 119909 ge 0

0 119909 lt 0(1)

where 119894119888and 119894119901(119901 = 0 119875 minus 1) denote the gray value of the

center pixel and gray value of the neighbor pixel on a circle ofradius 119877 respectively and 119875 is the number of the neighborsBilinear interpolation estimation method is used to estimatethe neighbors that do not lie exactly in the center of the pixelsLBPri119875119877

and LBPriu2119875119877

are rotation invariant of LBP and uniformrotation invariant of LBP respectively These two enhancedLBP operators are proposed by Ojala et al [23]

After completing the encoding step for any LBP opera-tors that is LBPri

119875119877and LBPriu2

119875119877 the histogram can be created

based on the following equation

119867(119896) =

119868

sum119894=0

119869

sum119895=0

119891 (LBP119875119877

(119894 119895) 119896) 119896 isin [0 119870]

119891 (119909 119910) = 1 119909 = 119910

0 otherwise

(2)

where119870 is the maximal LBP pattern value

22 Brief Review of LTP Tan and Triggs [27] presented anew texture operator which is more robust to noise The LBPis extended to 3-valued codes (minus1 0 1) Figure 4 shows anexample of LTP operatorThemathematical expression of theLTP can be described as follows

LTP119875119877

=

119875minus1

sum119901=0

2119901119904 (119894119901minus 119894119888) 119904 (119909) =

1 119909 ge 119905

0 minus119905 lt 119909 lt 119905

minus1 119909 lt minus119905

(3)


31 90 102

70 34 20

12 33 9

0

0

1 1

1

0

00

0

0

0

00 0

0

1

1

1

1

1

1

minus1

minus1minus1

Upper pattern

Lower pattern

t = 5

Threshold

[34 minus t 34 + t]

Binary01101000

Decimal104

Binary10000101

Decimal133






119901


119904119901= 119904 (119894119901minus 119894119888) 119898

119901=10038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816 (4)




CLBP S119875119877

=

119875minus1

sum119901=0

2119901119904 (119894119901minus 119894119888) 119904

119901=

1 119894119901ge 119894119888

0 119894119901lt 119894119888

(5)

CLBP M119875119877

=

119875minus1

sum119901=0

2119901119905 (119898119901 119888)

119905 (119898119901 119888) =

110038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816lt 119888

(6)




CLBP C119875119877

= 119905 (119894119888 119888119868) (7)


119868is



LBC119875119877

=

119875minus1

sum119901=0

119904 (119894119901minus 119894119888) 119904 (119909) =

1 119909 ge 0

0 119909 lt 0(8)


CLBC M119875119877

=

119875minus1

sum119901=0

119905 (119898119901 119888) 119905 (119898

119901 119888) =

110038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816lt 119888

CLBC C119875119877

= 119905 (119894119888 119888119868)

(9)

where 119894119888 119894119901 119888 and 119888



119875119877) is used to








31 90 102

70 34 20

12 33 9

0 1 1

1 0

000

LBP S = 01101000

(a)

3 56 68

36 14

22 1 25

0 1 1

1 0

000

Assume c = 29

LBP M = 01101000

(b)


52 90 102

70 34 20

12 40 9

1 1 1

1 0

0 1 0

34

ThresholdLBC code 5



119904upper119901

= 119904 (119894119901minus (119894119888+ 119905)) 119904

lower119901

= 119904 (119894119901minus (119894119888minus 119905))

119898upper119901

=10038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816 119898

lower119901

=10038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816

(10)







=

119875minus1

sum119901=0

2119901119904 (119894119901minus (119894119888+ 119905))

119904upper119901

= 1 119894119901ge 119894119888+ 119905

0 otherwise


=

119875minus1

sum119901=0

2119901119904 (119894119901minus (119894119888minus 119905))

119904lower119901

= 1 119894119901lt 119894119888minus 119905

0 otherwise(11)

The CLTP S119875119877



as follows

CLTP S119875119877

= [CLTP Supper119875119877


] (12)



the CLTP M119875119877


119901 and119898lower119901

as follows


=

119875minus1

sum119901=0

2119901119905 (119898

upper119901

119888)

119905 (119898upper119901

119888) =

110038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816lt 119888

(13)


=

119875minus1

sum119901=0

2119901119905 (119898

lower119901

119888)

119905 (119898lower119901

119888) =

110038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816lt 119888

(14)

CLTP M119875119877

= [CLTP Mupper119875119877


] (15)




119875119877can be



= 119905 (119894upper119888

119888119868)


= 119905 (119894lower119888

119888119868)

(16)


= 119894119888+ 119905 119894lower119888

= 119894119888minus 119905 and 119888








119894and119870 = 119896

119894where (119894 = 1 2 3 119861) can be


Dissimilarity1205942 (119867119870) =

119861

sum119894=1

(ℎ119894minus 119896119894)2

ℎ119894+ 119896119894

(17)












5 Conclusion




119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8






119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8










Acknowledgment


References











































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




31 90 102

70 34 20

12 33 9

0

0

1 1

1

0

00

0

0

0

00 0

0

1

1

1

1

1

1

minus1

minus1minus1

Upper pattern

Lower pattern

t = 5

Threshold

[34 minus t 34 + t]

Binary01101000

Decimal104

Binary10000101

Decimal133






119901


119904119901= 119904 (119894119901minus 119894119888) 119898

119901=10038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816 (4)




CLBP S119875119877

=

119875minus1

sum119901=0

2119901119904 (119894119901minus 119894119888) 119904

119901=

1 119894119901ge 119894119888

0 119894119901lt 119894119888

(5)

CLBP M119875119877

=

119875minus1

sum119901=0

2119901119905 (119898119901 119888)

119905 (119898119901 119888) =

110038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816lt 119888

(6)




CLBP C119875119877

= 119905 (119894119888 119888119868) (7)


119868is



LBC119875119877

=

119875minus1

sum119901=0

119904 (119894119901minus 119894119888) 119904 (119909) =

1 119909 ge 0

0 119909 lt 0(8)


CLBC M119875119877

=

119875minus1

sum119901=0

119905 (119898119901 119888) 119905 (119898

119901 119888) =

110038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus 119894119888

10038161003816100381610038161003816lt 119888

CLBC C119875119877

= 119905 (119894119888 119888119868)

(9)

where 119894119888 119894119901 119888 and 119888



119875119877) is used to








31 90 102

70 34 20

12 33 9

0 1 1

1 0

000

LBP S = 01101000

(a)

3 56 68

36 14

22 1 25

0 1 1

1 0

000

Assume c = 29

LBP M = 01101000

(b)


52 90 102

70 34 20

12 40 9

1 1 1

1 0

0 1 0

34

ThresholdLBC code 5



119904upper119901

= 119904 (119894119901minus (119894119888+ 119905)) 119904

lower119901

= 119904 (119894119901minus (119894119888minus 119905))

119898upper119901

=10038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816 119898

lower119901

=10038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816

(10)







=

119875minus1

sum119901=0

2119901119904 (119894119901minus (119894119888+ 119905))

119904upper119901

= 1 119894119901ge 119894119888+ 119905

0 otherwise


=

119875minus1

sum119901=0

2119901119904 (119894119901minus (119894119888minus 119905))

119904lower119901

= 1 119894119901lt 119894119888minus 119905

0 otherwise(11)

The CLTP S119875119877



as follows

CLTP S119875119877

= [CLTP Supper119875119877


] (12)



the CLTP M119875119877


119901 and119898lower119901

as follows


=

119875minus1

sum119901=0

2119901119905 (119898

upper119901

119888)

119905 (119898upper119901

119888) =

110038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816lt 119888

(13)


=

119875minus1

sum119901=0

2119901119905 (119898

lower119901

119888)

119905 (119898lower119901

119888) =

110038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816lt 119888

(14)

CLTP M119875119877

= [CLTP Mupper119875119877


] (15)




119875119877can be



= 119905 (119894upper119888

119888119868)


= 119905 (119894lower119888

119888119868)

(16)


= 119894119888+ 119905 119894lower119888

= 119894119888minus 119905 and 119888








119894and119870 = 119896

119894where (119894 = 1 2 3 119861) can be


Dissimilarity1205942 (119867119870) =

119861

sum119894=1

(ℎ119894minus 119896119894)2

ℎ119894+ 119896119894

(17)












5 Conclusion




119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8






119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8










Acknowledgment


References











































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




31 90 102

70 34 20

12 33 9

0 1 1

1 0

000

LBP S = 01101000

(a)

3 56 68

36 14

22 1 25

0 1 1

1 0

000

Assume c = 29

LBP M = 01101000

(b)


52 90 102

70 34 20

12 40 9

1 1 1

1 0

0 1 0

34

ThresholdLBC code 5



119904upper119901

= 119904 (119894119901minus (119894119888+ 119905)) 119904

lower119901

= 119904 (119894119901minus (119894119888minus 119905))

119898upper119901

=10038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816 119898

lower119901

=10038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816

(10)







=

119875minus1

sum119901=0

2119901119904 (119894119901minus (119894119888+ 119905))

119904upper119901

= 1 119894119901ge 119894119888+ 119905

0 otherwise


=

119875minus1

sum119901=0

2119901119904 (119894119901minus (119894119888minus 119905))

119904lower119901

= 1 119894119901lt 119894119888minus 119905

0 otherwise(11)

The CLTP S119875119877



as follows

CLTP S119875119877

= [CLTP Supper119875119877


] (12)



the CLTP M119875119877


119901 and119898lower119901

as follows


=

119875minus1

sum119901=0

2119901119905 (119898

upper119901

119888)

119905 (119898upper119901

119888) =

110038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus (119894119888+ 119905)

10038161003816100381610038161003816lt 119888

(13)


=

119875minus1

sum119901=0

2119901119905 (119898

lower119901

119888)

119905 (119898lower119901

119888) =

110038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816ge 119888

010038161003816100381610038161003816119894119901minus (119894119888minus 119905)

10038161003816100381610038161003816lt 119888

(14)

CLTP M119875119877

= [CLTP Mupper119875119877


] (15)




119875119877can be



= 119905 (119894upper119888

119888119868)


= 119905 (119894lower119888

119888119868)

(16)


= 119894119888+ 119905 119894lower119888

= 119894119888minus 119905 and 119888








119894and119870 = 119896

119894where (119894 = 1 2 3 119861) can be


Dissimilarity1205942 (119867119870) =

119861

sum119894=1

(ℎ119894minus 119896119894)2

ℎ119894+ 119896119894

(17)












5 Conclusion




119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8






119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8










Acknowledgment


References











































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





119894and119870 = 119896

119894where (119894 = 1 2 3 119861) can be


Dissimilarity1205942 (119867119870) =

119861

sum119894=1

(ℎ119894minus 119896119894)2

ℎ119894+ 119896119894

(17)












5 Conclusion




119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8






119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8










Acknowledgment


References











































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8






119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8










Acknowledgment


References











































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in





119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8



119877 = 3 119875 = 24 119877 = 2 119875 = 16 119877 = 1 119875 = 8










Acknowledgment


References











































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in











Acknowledgment


References











































Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in








Acknowledgment


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










Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



Documents

Research Article Completed Local Ternary Pattern for ...downloads.hindawi.com/journals/tswj/2014/373254.pdf · Ternary code ( 1 )1101 ( 1 )0( 1 ) F : LTP operator. where , , ,and