automatic characterization of the visual appearance of industrial

Hindawi Publishing CorporationAdvances in Optical TechnologiesVolume 2013 Article ID 503541 11 pageshttpdxdoiorg1011552013503541

Review ArticleAutomatic Characterization of the Visual Appearance ofIndustrial Materials through Colour and Texture AnalysisAn Overview of Methods and Applications

Elena Gonzaacutelez1 Francesco Bianconi2 Marcos X Aacutelvarez1 and Stefano A Saetta2

1 School of Industrial Engineering Universidade de Vigo Campus Universitario 36310 Vigo Spain2Department of Industrial Engineering Universita degli Studi di Perugia Via G Duranti 67 06125 Perugia Italy

Correspondence should be addressed to Elena Gonzalez elenauvigoes

Received 15 July 2013 Revised 16 September 2013 Accepted 16 September 2013

Academic Editor Pierre Chavel

Copyright copy 2013 Elena Gonzalez et alThis 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

We present an overview of methods and applications of automatic characterization of the appearance of materials through colourand texture analysis We propose a taxonomy based on three classes of methods (spectral spatial and hybrid) and discuss theirgeneral advantages and disadvantages For each class we present a set of methods that are computationally cheap and easy toimplement and that was proved to be reliable in many applications We put these methods in the context of typical industrialenvironments and provide examples of their application in the following tasks surface grading surface inspection and content-based image retrieval We emphasize the potential benefits that would come from a wide implementation of these methods suchas better product quality new services and higher customer satisfaction

1 Introduction

Computer vision has been a topic of intense research activityfor decades The wide availability of imaging devices aswell as the continuously increasing computing power hascontributed to the development of successful applications inmany areas remote sensing computer-aided diagnosis auto-matic surveillance crowd monitoring and food control arejust some examples [1] Manufacturing is also an importantsector that benefits from computer vision methods robotics[2] process automation [3] and quality control [4] are typicalapplications in this context Most of these applications sharethe same underlying problem the automatic characterizationof the visual appearance of the materials that one has to dealwith in the various domains

This is particularly true for products with high aestheticvalue [5 6] such as natural stone [7] ceramic [8] parquet[9] and fabric [10 11] to cite some In such cases it is thevisual appearance itself that largely determines the qualitymdashtherefore the pricemdashof the material The measurement ofvisual appearance can be considered a part of ldquosoft metrol-ogyrdquo the aim of which is the objective quantification of

the properties of materials that are determined by sensorialhuman response In an increasingly competitive worldwidemarket reliable and effective assessment of this feature is akey point to ensure high quality standard and the success ofa company

Traditionally the analysis of the visual appearancehas been accomplished by skilled operators This is alengthy tedious scarcely reproducible and largely subjectiveapproach [12] To overcome these issues many companiesare abandoning manual procedures and moving towardsautomated computer vision systems which in principleprovide higher quality standards better reproducibility andmore reliable product records These systems employ imageprocessing techniques to encode the visual appearance of aparticular product through the analysis of two main visualfeatures namely colour and texture even though othercharacteristics such as contrast and gloss can be consideredas well [5]

Intuitively colour and texture provide most of the infor-mation that we need to infer the material an object iscomposed of as well as to distinguish one material from

2 Advances in Optical Technologies

another Neuroimaging research has in fact confirmed thiscommon belief [13] Colour is the result of the interactionof three elements [14] an intrinsic characteristic of thematerial (reflectance spectrum) an accidental condition(illumination) and the response of the sensor (human eye orcamera) Texture a more elusive concept has been definedin many ways though none of the definitions given so far isentirely satisfactory Herein we will mention the definitiongiven by Davies [15] according to whom texture is ldquotheproperty of a surface that gives rise to the local variabilityin appearancerdquo where such variability may be either anintrinsic characteristic of the material or the consequence ofsurface roughness The concept of stationary texture whichmeans that the statistical characteristics of the texture are thesame everywhere [16] is also useful in this context In theremainder of the paper we implicitly assume that when weuse the word texture we refer to a stationary texture

Wepresent in this paper an overviewof colour and textureanalysis methods along with their applications in industryThe remainder of the manuscript is organized as followsSection 2 gives an up-to-date overview of the methods forautomatic characterization of the visual appearance andproposes a taxonomy to group them in a meaningful wayParticular attention is given to robustness against noisefactors such as variations in illumination and viewpointSection 3 deals with the main applications in industry withparticular emphasis on grading surface inspection andcontent-based image retrieval (CBIR) Concluding remarksare reported in Section 4 along with a discussion on futureresearch directions and potentially innovative applications

2 Methods

Originally approaches to automatic characterization of thevisual appearance of materials considered colour and textureseparately In the last decade however an increasing interesthas been devoted to integrating these two sources of data intoa unifying model Strong evidence has in fact supported thisidea and confirmed that incorporating colour into the texturemodel has beneficial effects in many applications [17 18]Theclassification scheme that we propose in this paper reflectsthis idea namely that themethods can be based on one of thetwo features separately (either colour or texture) or considerthe two of them jointly

Based on these considerations itmakes sense to define thefollowing three main categories spectral spatial and hybridmethods [19] The first group is based on colour only therelative spatial distribution of the pixel values is not taken intoaccount The second group encloses texture-based methodsthey are called spatial methods since they take into accountthe relative variation of pixel intensity (grayscale values) inthe spatial domain but discards colour altogether (images arepreviously converted into grayscale) Finally those methodsthat consider colour and texture jointly are referred to ashybrid methods Before describing the three groups in detailwe will first discuss three points that apply to each method ingeneral these are dimensionality robustness to noise factorsand parameter tuning

Dimensionality refers to the number of parameters amethod relies upon Whichever the approach we use tocharacterize the visual appearance of a material this will berepresented through a finite set of quantitative parameterswhich are usually referred to as the feature vectorThe appear-ance of a material can therefore be regarded as a point in an119899-dimensional space where 119899 is the number of parametersIdeally a good method should provide high discriminationaccuracy with as few parameters as possible Computation infact becomes costly as dimensionality increases representinga serious limitation for applications that require real-timeprocessing Furthermore long feature vectors are likely todegrade the performance of a classifier as a consequence ofthe well-known phenomenon of ldquocurse of dimensionalityrdquo[20]

Noise factors represent uncontrollable sources of uncer-tainty that commonly occur in nonideal conditions Inreal working environments it can be quite complicated tomaintain steady and controlled imaging conditions Mostcommonly one has to deal with variable conditions such aschanges in rotation scale illumination and so forth It can bequite difficult to compensate for such sources of noise Someof them are likely to degrade the performance of a methodso strongly as to make it virtually useless To illustrate thisconcept Figure 1 shows how the appearance of a material canbe affected when exposed to different sources of noise

The effects of changes in illumination are reported inFigures 1(a) and 1(b) Changes in translation scale rotationandor viewpoint are shown in Figures 1(c) 1(d) 1(e) and1(f) As we discuss in the following subsections each classof methods is sensitive to one or more sources of noise Itis therefore the designerrsquos responsibility to select the methodwhich is most appropriate to the specific application domainconsidering the potential noise factors that can arise in theirspecific context

Finally it is important to mention that virtually anyspectral spatial or hybrid image descriptor depends onone or more parametersmdashsee Table 1 for a roundup of themethods presented here and their parameters In any specificapplication the performance of a method may vary greatlydepending on the setup of such parametersWhen it comes toimplementing industrial systems it is advisable to stick to thevalues suggested in the literature as a first step then to furtherimprove the results it is possible to optimise the parametersrsquovalues in a supervised fashion adaptive to thematerial to theimage kind but also to the experience of the user

In the following subsections we analyse spectral spatialand hybrid approaches to colour texture analysis and discusstheir pros and cons For each class we present a set ofrepresentativemethods and their use in three different indus-trial applications grading surface inspection and CBIR InTable 2 we summarise the robustness of each method todifferent noise factors (ie changes in illumination inten-sity viewpoint rotation and scale) and report a qualitativeappraisal of the computational time The methods includedin the experiments have been selected on the basis ofclassification accuracy computational efficiency and easinessof implementation

Advances in Optical Technologies 3

(a) (b) (c)

(d) (e) (f)

Figure 1 (a) Frontal view of a covered board and its appearance under noise factors such as changes in (b) illumination (c) translation(d) scale (e) rotation and (f) viewpoint

Table 1 Method and related parameters

Class Method Parameters

Spectral

R G B means NoneColour histogram Number of quantization levels

Chromaticity moments Number of quantization levelsColour centiles Number of percentiles

Spatial

GLCM Distance co-occurrence directionsLBPri Radius number of pixelsCCRri Radius number of pixels and thresholding method

Gabor filtersNumber of frequencies number of orientations maximumfrequency width and length of the Gaussian envelope and

frequency ratio

Hybrid

OCLBPri Radius number of pixelsIntegrative CM Distance co-occurrence directions

Ranklets Radius number of pixelsLBPri + colour centiles Radius number of pixels and number of percentiles

21 Spectral Methods Spectral methods consider the colourcontent of the image regardless of its spatial distributionThis is usually represented through RGB triplets since thisis the format used by most imaging devices Hyperspectralimaging is also gaining importance in practical applicationsfor a review ofmethods the interested reader is referred to thework of Chang [21] Herein we limit our discussion to the tra-ditional and well-established domain of trichromatic imagesIt is worth recalling at this point some basic facts aboutcolour spaces We know that they can be device-independent(colorimetric) or device-dependent The first group includesany colour space that can be converted into XYZ withoutadditional information In principle device-independent dataare absolute values that do not depend on the imagingsystem They also carry information about the illuminantunder which the measurement is taken (eg D65 D50etc) In contrast device-dependent colour spaces encode

device-specific data Consequently colour data acquired withtwo different devices are unrelated and no information isconveyed about the illuminant

The problem of choosing the best colour space for visualsimilarity has been debated at length in the literature [1822 23] As for robustness against changes in illuminationit is clear that in principle device-independent data are thebest option The problem however is that standard imagingdevices do not provide this kind of datamdashwhich need tobe estimated through calibration Regarding accuracy it hasbeen suggested that linear spaces (eg CIE Lab) shouldprovide the best results for distance in these spaces correlateswith dissimilarity in the human vision system Experimentalresults however have not confirmed this hypothesis theyreport on the contrary rather inconclusive and fairly con-tradictory results [24 25]


Table 2 Robustness to noise factors and computing time A tick indicates that the method is robust to the corresponding noise factor Lowmedium and high computing time are indicated as ∙ ∙∙ and ∙ ∙ ∙ respectively

Class MethodNoise factor

Computing timeChange inillumination intensity

View pointchange Rotation Scale

Spectral

R G B means ∙

Colour histogram ∙∙

Chromaticity moments ∙∙

Colour centiles ∙

Spatial

GLCM ∙∙

LBPri81 ∙

CCRri81 ∙∙

Gabor filters ∙∙

Hybrid

OCLBPri81 ∙∙

Integrative CM ∙∙

Ranklets ∙ ∙ ∙

LBPri81 + colour centiles ∙

We subdivide spectral methods into two subgroupsmethods based on colour statistics and methods based oncolour histograms The methods of the first group considervarious combinations of global statistical parameters (suchas mean value standard deviation median etc) which arecomputed directly from the colour data These are alsoreferred to as soft colour descriptors [26] Within this groupit has been shown that parameters as simple as the averagevalues of each R G and B channel can discriminate quitewell among different materials Kukkonen et al [8] reporteda successful application of this method in surface gradingof ceramic tiles In the same domain of application Lopezet al [26] proved the effectiveness of various combinations ofsoft colour descriptors such as mean standard deviation andmoments Niskanen et al proposed the use of colour centiles(values that divide the cumulative probability distribution ofeach colour channel into the percentage required) for defectdetection in parquet slabs [27]

The second group is based on the concept of colour his-togramThis is an approximated estimation of the probabilityof occurrence of each colour in an image The number ofcomponents of the colour histogram is determined througha suitable quantization of the colour space The progenitorof this group is the 3D colour histogram [28] This isobtained by dividing the colour space into cells of equalvolume This method is still valid today but presents thedisadvantage of a high dimensionality Modified versionshave been proposed with the aim of avoiding this problemAmong these it is worth mentioning Paschosrsquo chromaticitymoments [22] which discard the intensity channel andconsider the two-dimensional histogram arising from thetwo chromatic channels only The colour images are firstconverted into the 119909119910119884 spacemdasha process which requirescolour calibrationmdashand then the Y (intensity) channel isdiscarded The resulting two-dimensional distribution overthe 119909119910 plane is characterized through a set of moments

We conclude this subsection with some remarks aboutspectral methods as a whole It is important to point outthat due to the fact that this family of techniques considersthe colour content of the image regardless of the spatialdistribution any spatial information is lost A positive out-come of that is a fairly good robustness against translationrotation and changes of the viewing angle (see Table 2)A negative one is that dissimilar images may have similarcolour distributions and therefore these methods can fail insome cases To illustrate this concept we captured an imageof a granite slab (Figure 2(a)) and rearranged the spatialdistribution of the pixels (Figure 2(b)) Both images haveexactly the same colour content but their appearance is quitedifferent

In summary spectral methods are highly recommendedin presence of steady illumination conditions In such situa-tions they can compensate very well for changes in viewpointillumination and scale By contrast they should be avoidedwhen illumination conditions are variable or unknownFor quantitative performance analyses on spectral methodsapplied to wood and ceramics the interested reader may finduseful data in [8 9 24 26]

22 Spatial Methods Spatial methods are based on puretexture They capture the visual appearance of a materialthrough features extracted from grayscale images thereforediscarding colour information Research on this area has beenintense for many years Various attempts to categorize themethods of this group have been proposed in the past [29ndash31] Such classification schemes however have proved ratherunsatisfactory due to the high degree of overlap that existsamong many methods In an effort to solve this problem Xieand Mirmehdi have recently proposed a fuzzy categorization[32] In their scheme a method can belong to one or more ofthe following four classes statistical structural model-basedand signal processing-based Statistical approaches considerthe spatial distribution of pixel values Structural methods


(a) (b)

Figure 2 (a) Image of a granite slab (b) Synthetically generated image obtained by rearranging the pixels of image (a)

describe textures in terms of the spatial arrangement ofcertain elementary patterns which are usually referred toas texels or textons Model-based methods characterize thetexture through a set of parameters of a predefined mathe-matical model Finally the signal processing-based methodstypically employ a bank of filters to analyse texture at differentfrequencies and orientations Herein we selected four classicmethods that are easy to implement and proved to be effectivein many applications gray level co-occurrence matriceslocal binary patterns coordinated clusters representationand Gabor filters

Gray level co-occurrence matrices (GLCM) are amongthe most common texture descriptors [33] They are alsoimportant from a historical perspective as they are one ofthe methods that pioneered texture analysis This approachbelongs to the statistical group It is based on the bidi-mensional joint probability of the grayscale values of pairsof pixels separated by a predefined displacement vectorEach displacement generates one co-occurrencematrix fromwhich a set of statistical parameters are computed Hereinwe considered the following contrast correlation energyentropy and homogeneity Most commonly the set of dis-placements includes one-pixel shifts along four directions0∘ 45∘ 90∘ and 135∘ This is the setting adopted in ourimplementationThe matrices corresponding to the four dis-placements are averaged for rotation-invariance Invarianceto illumination changes can be obtained by using statisticalparameters that are invariant to illumination intensity suchas energy entropy contrast correlation and homogeneity

Local binary patterns (LBP) [34] and coordinated clustersrepresentation (CCR) [35] are both statistical and struc-tural They characterize textures through the probability ofoccurrence of certain elemental patterns that can occur ina neighbourhood of predefined shape and size Experimentshave shown that good results can be obtained using aneighbourhood as small as a 3 times 3 window Both LBP andCCR consider the probability of occurrence of the binarypatterns that result from each position of the neighbourhoodas it slides by one-pixel steps along the image Binarizationis based on thresholding this is local in the case of LBP(threshold is the value of the central pixel) and global in thecase of CCR (threshold is computed from the whole imagethrough isentropic partition of the grayscale histogram)Thisgives 28 and 29 possible binary patterns in the two cases (LBPand CCR resp) Rotationally invariant versions of the two

methods can be easily obtained considering an interpolatedcircular neighbourhood [34] These two versionsmdashwhich arethe ones used hereinmdashare referred to in the literature asLBPri81

and CCRri81

[36]Gabor filters [37] have played a significant role in texture

analysis and are ubiquitous in many applications The reasonfor the success of the method perhaps lies in its capability ofmimicking the human vision system [38] The approach isbased on multichannel frequency- and orientation-selectiveanalysis of the image The implementation therefore consistsin designing a bank of filters and selecting a proper set ofparameters for each filter [39] Texture features are typicallythe mean and the standard deviation of the absolute value ofthe transformed images Invariance against rotation can beeasily achieved through DFT normalization

Compared with spectral methods spatial methods showrather opposed characteristics Whereas the former areinvariant against changes in rotation viewpoint and scalethe latter are not They are in fact quite sensitive to changesin viewpoint scale and rotation Invariance against rotationcan be easily introduced in some spatial methods (themethods considered herein are all rotationally invariant) butinvariance to viewpoint and scale is much more difficult toachieve In contrast whereas spectral methods are scarcelyresilient to illumination changes structural methods aremore robust in this sense Some of them are intrinsicallyinvariant to changes in illumination intensity (eg LBP)in the others a change in illumination intensity can beeasily compensated for In summary spatial methods arerecommendable when we are interested in discriminatingmaterialsrsquo appearance on the basis of local changes that istexture Hence they are appealing for those materials with aclear local structure such as woven fabric wood granite andso forth by contrast they are scarcely descriptive of materialswith ldquoflatrdquo appearanceOperating conditions can tolerate onlyslight changes in viewpoint and scale whereas changes inillumination intensity and rotation are acceptable when theproper methods are used (see Table 2)

23 Hybrid Methods In an attempt to exploit structural andspectral information jointly researchers have devoted sig-nificant attention to the development of combined (hybrid)methods Experiments have in fact confirmed that combinedapproaches can provide better performance when comparedwith methods that consider texture and colour separately


[18] Approaches that consider the two sources of data canrely on one of the following strategies considering colourand texture separately and concatenating the resulting featurevector extracting texture features from each colour channelseparately (intra-channel features) and extracting texturefeatures from couples of feature channels (inter-channelfeatures) There is no general consensus at present aboutwhich of these strategies is the best [26] In this subsection wereport four methods that have proved reliable and effective ina number of applications Integrative co-occurrencematricesopponent colour local binary patterns colour ranklets andlocal binary patterns + colour percentiles

GCLM can be easily extended to colour images consid-ering both the co-occurrence probability within one colourchannel (intra-channel features) and between couples ofcolour channels (inter-channel features) The latter are usu-ally extracted from R-G R-B and G-B couples as proposedin the implementation of Palm [40] Texture features areextracted as in GCLM The method is usually referred to asIntegrative co-occurrence matrices (integrative CM)

Local binary patterns have also been extended to colourimages in a similarwayThemethod is referred to as opponentcolour local binary patterns (OCLBP) In this model intra-channel features are obtained by applying the LBP operator toeach channel separately Inter-channel features are computedfor pairs of colour channels (ie R-G R-B andG-B) by takingthe neighbourhood from one channel and the threshold fromthe other channel [41] The method considered herein makesuse of the rotationally invariant version of LBP LBPri

81

andand is therefore indicated as OCLBPri

81

In a similar way ranklets a nonparametric local measure

of relative intensity based on the ranklet transform have beenextended to colour textures by taking both intra- and inter-channel features in the RGB space [42] Since both colourranklets and OCLBPri are based on relative differences orratios between colour values in the R G and B channelsthey are theoretically invariant to changes in illuminationintensity

Finally we mention an approach where texture andcolour are computed separately and the resulting features aremerged This method considers LBPri

81

as texture descriptorand colour centiles as colour descriptors Successful applica-tions of this method have been reported in surface detectionof parquet slabs [27]

In Sections 21 and 22 we have set into evidence thatspectral and spatial methods capture different and comple-mentary aspects of materialsrsquo appearance spectral methodsare able to discriminate among material with similar texturebut different colour whereas spatial methods can distinguishbetween similar colours and different textures It makestherefore sense to combine the two concepts into jointdescriptors However it has been pointed out that the gainin accuracy that one obtains with hybrid methods pairs witha significant loss in robustness [43] In fact whereas spectraland spatial methods alone have some interesting invariancefeatures which have been mentioned in the preceding sub-sections these are lost on hybrid methods (see Table 2)Any type of change in illumination viewpoint rotation and

scale can in principle affect these methods and result in asignificant loss of performance Quantitative performanceanalyses of hybrid methods applied to granite and wood canbe found in [7 27]

3 Industrial Applications

Automatic characterization of the visual appearance of mate-rials has many interesting applications in industry In thispaper we are particularly concerned with the following mainclasses of problems grading surface inspection and content-based image retrieval The aim of this section is to providethe reader with a description of these scenarios as wellas to present practical examples of application in the threecontexts

Grading is a process where products are grouped intolots based on the criterion of ldquosimilar appearancerdquo [24]This means deciding which category or grade (among aset of predefined classes) each material belongs to [30]Grading applications are related to stages of the supply chainsuch as ordering delivering and warehousing This processimplies two major steps (1) extraction of visual featuresand (2) assignment of labels In the preceding section wepresented a set of methods that can be used in the firststep The second step is usually referred to as a classificationproblem which can be either supervised or unsupervisedIt is beyond the scope of this paper to describe in detailthese concepts and related methods The interested readerwill find a comprehensive reference in the classic text of Dudaet al [20] In practical applications supervised classificationconsists in assigning a class label to a product among a set ofpredefined classes each possible class being represented byone or more training samples In contrast in unsupervisedclassification we are given a set of products to be grouped intoa predefined number of classes but no training samples aregiven in this case Unsupervised classification is also referredto as clustering

Surface inspection is concerned with the detection ofvisible surface defects such as stains cracks veins knotsand so forth An interesting and up-to-date review of recentadvances in this field can be found in [44] Surface inspec-tion is based on image segmentation a process where animage of the material under control is split into regionswith homogeneous visual properties Segmentation has beenextensively studied by the computer vision community andmany approaches are described in the literature They rangefrom simple pixel-by-pixel classification to more involvedmethods such as region growing

Content-based image retrieval is the process of searchinglarge datasets based on visual content rather than metadatasuch as keywords tags and so forth For a review of methodsreaders may find the work of Vassilieva useful [45] Potentialapplications of CBIR in industry are appealing though notcompletely exploited so far An interesting use for instance isin the problem of finding thematerial that is most similar to aquery sampleWe can imagine for example a situationwherea customer needs to replace some broken pieces of graniteceramics and so forth or has to enlarge a facade or floorwhich has previously been covered with material of a certain


Table 3 Results of the grading experiment

Class Method Dimension DatabaseWood Ceramic Textile Granite

Spectral

R G B means 3 946 853 993 890Colour histogram 2048 948 846 994 954

Chromaticity moments 6 887 795 1000 918Colour centiles 15 982 870 992 892

Spatial

GLCM 5 548 466 702 595LBPri

81 36 964 840 945 806CCRri

81 72 844 698 894 796Gabor filters 32 991 818 906 956

Hybrid

OCLBPri81 216 980 876 1000 962

Integrative CM 30 767 748 915 857Ranklets 216 989 858 1000 975

LBPri81 + colour centiles 55 984 892 1000 936

(a) Wood database (b) Textile database

(c) Ceramic database (d) Granite database

Figure 3 Image databases

type In such cases the customer will request a materialldquosimilar tordquo his (the query sample) A CBIR systemmdashperhapsdistributed over a network of potential providersmdashwould beof great help in such a case and would foster cooperationamong companies in a global marketplace

31 Grading In this section we present an example ofsurface grading applied to four different types of indus-trial materials wood ceramics textile and granite Thefirst three databases (Figures 3(a) 3(b) and 3(c)) arecomposed of 30 classes eachmdashone image per class Theimages have been picked from a public internet reposi-tory (httpwwwarchibaseplanetcom) and the resolutionis 400 times 400 pixels In order to get more examples of eachmaterial the images have been subdivided into four nonover-lapping subimages The fourth dataset (Figure 3(d)) contains12 granite classes with four images for each classThe graniteimages (httpdismacdiiunipgitmm) have been acquiredin our laboratory under controlled illumination conditionsusing an illuminator (Monster DOME Light 1825) and adigital camera (Samsung S850) Image resolution is 544times544pixels and no further subdivision is performed in this case

To assess the performance of the descriptors presented inSection 2 we carried out a supervised classification experi-ment First each dataset is randomly split into two subsets

one for training (training set) and the other for classification(validation set) The subdivision guarantees that half of theimages of each class are used for training and half forvalidation (this procedure is known as stratified sampling)Second each image of the validation set is classified throughthe nearest neighbour rule (1-NN) and 119871

1distance [20]

Finally classification accuracy is estimated as the percent-age of images of the validation set that have been classifiedcorrectly To make the results stable the subdivision intotraining and validation set is repeated 100 times and the finalaccuracy is the average value over the 100 problems Thegeneral expression to compute the classification accuracy is

119886 =

1

119875

119875

sum

119901=1

119862119901

119881119901

(1)

where 119886 is the classification accuracy 119875 the total number ofproblems (119875 = 100 in this case) 119862

119901the number of correctly

classified samples and 119881119901the number of the validation

samples in the 119901th problemThe results of the experiments are summarized in Table 3

The first two columns report the class and name of eachmethod the third column the dimension of the feature spaceand the following columns the average accuracy over the fourdatasets


Original image Ground truth

Spectral Spatial Hybrid

Segmentation based on R G B values

Segmentation Segmentation

Fabric

Paper

Wood

colour centiles

A = 9891 A = 9865 A = 9926

A = 9532 A = 9856 A = 9796

A = 9929 A = 9530 A = 9784

based on based on +LBPri81

LBPri81

Figure 4 Results of the surface inspection experiment

In general the results are fairly good and confirm theeffectiveness of the considered methods for grading Amongthe spatial methods LBPri

81

and Gabor filters show the bestaccuracy Spectral methods perform quite well too it is worthmentioning that a method as simple as the mean values ofeach R G and B channel provides rather a good accuracyHybrid methods on average are the most effective as onewould expect

32 Surface Inspection In this section we describe a typicalsurface inspection task consisting of detecting defective areason the inspected surfaceThree images of defective specimensof fabric paper and wood (see Figure 4) have been acquiredin our laboratory under controlled illumination conditionsusing the setup described in Section 31The task is about seg-menting the image into defective and non-defective zones Inthe experiment we avoided complicated segmentingmethodssuch as region growing or morphological analysis and stuckto the simple and easy-to-implement per pixel classificationWe considered one method for each of the spectral spatialand hybrid groups presented in Section 2 namely R G andB values LBPri

81

and LBPri81

+ colour centiles Segmentationhas been performed through supervised classification basedon the Naıve Bayes algorithm [20] with a normal probabilitydensity kernel For each specimen approximately 5 of thewhole area have been used for training and the remaining95 for validation The ldquotruerdquo position and extension of the

defects (ground truth) have been manually determined byhuman experts

As a measure of accuracy we considered the sum of thepercentage of foreground pixels (ie defective areas) cor-rectly classified as foreground and that of background pixels(ie non-defective areas) correctly classified as backgroundThis parameter gives an overall estimate of the effectivenessof the segmentation process In formulas we have

119860 =

1003816100381610038161003816119861 cap 119861119879

1003816100381610038161003816

|119868|

+

1003816100381610038161003816119865 cap 119865119879

1003816100381610038161003816

|119868|

(2)

where 119860 is the overall accuracy 119868 the whole image 119861 and 119865the background and foreground produced by the automaticsegmentation procedure 119861

119879and 119865

119879the ldquotruerdquo background

The segmentation results are summarised in Figure 4 Theresults show that for each problem there is at least onemethod that produces good segmentation results with 119860 gt98 5 Conversely not all the methods perform well in eachproblem suggesting that the selection of the right descriptoris a domain-specific problem which needs to be consideredwith extreme care Note that the results presented in Figure 4could be further enhanced through some morphologicalpostprocessing steps such as region growing morphologicalfiltering and so forth

33 Content-Based Image Retrieval Finally we present anapplication of content-based image retrieval To illustrate this


(a) (b)

Figure 5 (a) User interface for inserting new images in the database (b) User interface image retrieval

Figure 6 Database of 30 types of wood used in the retrievalexperiment

application we use a prototype software developed by theauthors (Figure 5)

In the example presented in this section we considereddifferent types of wood (parquet) The first step consists increating a database of wood images In this phase the userselects a set of images and uploads each image to the databaseWhen an image is uploaded by the user (Figure 5(a)) thesystem computes the visual characteristics of the image usingthe descriptors discussed in Section 2 (the user can chooseone ormore)The visual characteristics are then storedwithinthe imageThis process could be made automatically throughan appropriate online acquisition and recording system Inthe example shown in this section we created a database of 30types of wood (see Figure 6)

Once the database is created the user can insert a queryimage and perform content-based retrieval In this stage theuser selects the visual characteristic that he wants to use forthe retrieval and the distance type The system presents theuser with the four most similar images in the database indescending similarity order from left to right (Figure 5(b))The results of a retrieval experiment are reported in Figure 7Here we have a query image on the left and the retrievedimages on the right Each row corresponds to a differentdescriptor

It is interesting to comment on the result providedby the three methods First notice that the query image

presents a slight texture formed by approximately verticallines We can see that the first three images retrievedthrough LBPri

81

actually present a texture very similar tothat of the query imagersquos as one would expect On theother hand the first two images returned by colour centilesare very similar in colour to the query image but havevery little texture Finally the combination of LBPri

81

andcolour centiles provides a balanced result between colour andtexture

4 Conclusions

In this paper we presented an overview of methods for auto-matic characterization of the visual appearance of materialsand illustrated a set of applications in industry The aim ofthe paper was to provide both researchers and practitionerswith a survey of methods and applications in an area whereresearch interest is currently high The implementation ofthe methods presented in this paper may in fact lead tobetter products as well as new services This is particularlyinteresting for products where visual appearance plays acrucial role marble granite ceramic parquet leather fab-ric and paper are just some examples In an increasinglycompetitive global market effective measurement recordingand retrieval of visual features are key points to ensurehigh quality standards The benefits expected from a wideradoption of these methods include better product qualityfewer faulty products and higher customer satisfaction Inthe near future emerging techniques such as RFID arelikely to play an important role in this context It is easyto imagine a scenario where the visual characteristics of aproduct are stored together with the product in RFID tagsin some sort of product identity card This would bringsignificant improvement to warehouse management and thesupply chain

Conflict of Interests

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


Query image Retrieved images Method Class

Colour centiles Spectral

Spatial

Hybrid+ colour centiles

LBPri81

LBPri81

Figure 7 Results of the CBIR experiment

Acknowledgments

This work was partially supported by the Spanish Govern-mentwithin Project no CTM2010-16573 and by the EuropeanCommission within Projects no Life09-ENVFI000568 andno LIFE12-ENVIT000411

References

[1] E R DaviesMachine Vision Theory Algorithms PracticalitiesElsevier Morgan Kaufmann 2005

[2] Y H Jen Z Taha and L J Vui ldquoVR-Based robot programmingand simulation system for an industrial robotrdquo InternationalJournal of Industrial Engineering vol 15 no 3 pp 314ndash3222008

[3] Y Kwon ldquoWeb-Enabled vision guided robotic tracking withinthe framework of e-manufacturingrdquo International Journal ofIndustrial Engineering vol 15 no 3 pp 323ndash329 2008

[4] C-C Wang and B C Jiang ldquoPCB solder joint defects detectionand classification using machine visionrdquo International Journalof Industrial Engineering vol 8 no 4 pp 359ndash369 2001

[5] CIE ldquoA framework for the measurement of visual appearancerdquoTech Rep CIE 1752006 International Commission on Illumi-nation 2006

[6] C Eugene ldquoMeasurement of ldquototal visual appearancerdquo a CIEchallenge of soft metrologyrdquo in Proceedings of the 12th IMEKOTC1-TC7 Joint Symposium onMan Science ampMeasurement pp61ndash65 Annecy (France) 2008

[7] F Bianconi E Gonzalez A Fernandez and S A SaettaldquoAutomatic classification of granite tiles through colour andtexture featuresrdquo Expert Systems with Applications vol 39 no12 pp 11212ndash11218 2012

[8] S Kukkonen H Kalviainen and J Parkkinen ldquoColor featuresfor quality control in ceramic tile industryrdquoOptical Engineeringvol 40 no 2 pp 170ndash177 2001

[9] F Bianconi A Fernandez EGonzalez and SA Saetta ldquoPerfor-mance analysis of colour descriptors for parquet sortingrdquoExpertSystems with Applications vol 40 no 5 pp 1636ndash1644 2013

[10] P Bison A Bortolin G Cadelano G Ferrarini and EGrinzatoldquoEmissivitymeasurement of semitransparent textilesrdquoAdvancesin Optical Technologies vol 2012 Article ID 373926 5 pages2012

[11] M Carfagni R Furferi and L Governi ldquoA real-time machine-vision system for monitoring the textile raising processrdquo Com-puters in Industry vol 56 no 8-9 pp 831ndash842 2005

[12] C Boukouvalas J Kittler R Marik M Mirmehdi and MPetrou ldquoCeramic tile inspection for colour and structuraldefectsrdquo in Proceedings of the AdvancedMaterials and ProcessingTechnologies pp 390ndash399 Dublin Ireland 1995

[13] E H Adelson ldquoOn seeing stuff the perception of materialsby humans and machinesrdquo in Human Vision and ElectronicImaging VI B E Rogowitz and T N Pappas Eds vol 4299of Proceedings of the SPIE pp 1ndash12 January 2001

[14] G Wyszecki and W Styles Color Science Concepts and Meth-ods Quantitative Data and Formulae Wiley-Interscience 2ndedition 1982

[15] E R Davies ldquoIntroduction to texture analysisrdquo in Handbook ofTexture AnalysisMMirmehdi X Xie and J Suri Eds pp 1ndash31Imperial College Press 2008

[16] M Petrou and P Garcıa Sevilla Image Processing Dealing withTexture Wiley-Interscience 2006

[17] M C Ionita and P Corcoran ldquoBenefits of using decorrelatedcolor information for face segmentationtrackingrdquo Advances inOptical Technologies vol 2008 Article ID 583687 8 pages 2008

[18] A Drimbarean and P F Whelan ldquoExperiments in colourtexture analysisrdquo Pattern Recognition Letters vol 22 no 10 pp1161ndash1167 2001

[19] E Gonzalez F Bianconi and A Fernandez ldquoA comparativereview of colour features for content-based image retrievalrdquoAnales de Ingenierıa Grafica vol 21 pp 7ndash14 2010

[20] R O Duda P E Hart and D G Stork Pattern ClassificationWiley-Interscience 2nd edition 2001

[21] C-I Chang Hyperspectral Imaging Techniques for SpectralDetection and Classification Plenum 2003

[22] G Paschos ldquoFast color texture recognition using chromaticitymomentsrdquoPatternRecognition Letters vol 21 no 9 pp 837ndash8412000

[23] E Valencia and M S Millan ldquoColor image quality in presen-tation softwarerdquo Advances in Optical Technologies vol 2008Article ID 417976 6 pages 2008

[24] F Lopez J Miguel Valiente J Manuel Prats and A FerrerldquoPerformance evaluation of soft color texture descriptors forsurface grading using experimental design and logistic regres-sionrdquo Pattern Recognition vol 41 no 5 pp 1761ndash1772 2008

[25] C Vertan and N Boujemaa ldquoColor texture classification bynormalized color space representationrdquo in Proceedings of the15th International Conference on Pattern Recognition (ICPR rsquo00)vol 3 pp 580ndash583 Barcelona Spain 2000

[26] F Lopez J M Valiente R Baldrich and M Vanrell ldquoFastsurface grading using color statistics in the CIE lab spacerdquo in


Proceedings of the 2nd IberianConference on Pattern Recognitionand Image Analysis (IbPRIA rsquo05) vol 3523 of Lecture Notes inComputer Science pp 666ndash673 June 2005

[27] M Niskanen O Silven and H Kauppinen ldquoColor and texturebased wood inspection with non-supervised clusteringrdquo inProceedings of the 12th Scandivanian Conference on ImageAnalysis (SCIA rsquo01) pp 336ndash342 Bergen Norway 2001

[28] M J Swain and D H Ballard ldquoColor indexingrdquo InternationalJournal of Computer Vision vol 7 no 1 pp 11ndash32 1991

[29] M Sonka V Hlavac and R Boyle Image Processing Analysisand Machine Vision Thomson 3rd edition 2005

[30] M Tuceryan and A K Jain ldquoTexture analysisrdquo in Handbookof Pattern Recognition and Computer Vision C H Chen LF Pau and P S P Wang Eds pp 207ndash248 World ScientificPublishing 2nd edition 1998

[31] C-MWu and Y-C Chen ldquoStatistical feature matrix for textureanalysisrdquo CVGIP Graphical Models and Image Processing vol54 no 5 pp 407ndash419 1992

[32] X Xie and M Mirmehdi ldquoA galaxy of texture featuresrdquo inHandbook of Texture Analysis M Mirmehdi X Xie and J SuriEds pp 375ndash406 Imperial College Press 2008

[33] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[34] 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

[35] E V Kurmyshev and R E Sanchez-Yanez ldquoComparativeexperimentwith colour texture classifiers using theCCR featurespacerdquo Pattern Recognition Letters vol 26 no 9 pp 1346ndash13532005

[36] A Fernandez O Ghita E Gonzalez F Bianconi and P FWhelan ldquoEvaluation of robustness against rotation of LBP CCRand ILBP features in granite texture classificationrdquo MachineVision and Applications vol 22 no 6 pp 913ndash926 2011

[37] B S Manjunath ldquoTexture features for browsing and retrieval ofimage datardquo IEEE Transactions on Pattern Analysis andMachineIntelligence vol 18 no 8 pp 837ndash842 1996

[38] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[39] F Bianconi and A Fernandez ldquoEvaluation of the effects ofGabor filter parameters on texture classificationrdquoPattern Recog-nition vol 40 no 12 pp 3325ndash3335 2007

[40] C Palm ldquoColor texture classification by integrative co-occurrence matricesrdquo Pattern Recognition vol 37 no 5 pp965ndash976 2004

[41] T Maenpaa and M Pietikainen ldquoTexture analysis with localbinary patternsrdquo in Handbook of Pattern Recognition andComputer Vision C H Chen and P S P Wang Eds pp 197ndash216 World Scientific Publishing 3rd edition 2005

[42] F Bianconi A Fernandez E Gonzalez and J Armesto ldquoRobustcolor texture features based on ranklets and discrete Fouriertransformrdquo Journal of Electronic Imaging vol 18 no 4 ArticleID 043012 2009

[43] T Maenpaa and M Pietikainen ldquoClassification with color andtexture jointly or separatelyrdquo Pattern Recognition vol 37 no8 pp 1629ndash1640 2004

[44] X Xie ldquoA review of recent advances in surface defect detectionusing texture analysis techniquesrdquo Electronic Letters on Com-puter Vision and Image Analysis vol 7 pp 1ndash22 2008

[45] N S Vassilieva ldquoContent-based image retrieval methodsrdquoProgramming and Computer Software vol 35 no 3 pp 158ndash1802009

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



another Neuroimaging research has in fact confirmed thiscommon belief [13] Colour is the result of the interactionof three elements [14] an intrinsic characteristic of thematerial (reflectance spectrum) an accidental condition(illumination) and the response of the sensor (human eye orcamera) Texture a more elusive concept has been definedin many ways though none of the definitions given so far isentirely satisfactory Herein we will mention the definitiongiven by Davies [15] according to whom texture is ldquotheproperty of a surface that gives rise to the local variabilityin appearancerdquo where such variability may be either anintrinsic characteristic of the material or the consequence ofsurface roughness The concept of stationary texture whichmeans that the statistical characteristics of the texture are thesame everywhere [16] is also useful in this context In theremainder of the paper we implicitly assume that when weuse the word texture we refer to a stationary texture

Wepresent in this paper an overviewof colour and textureanalysis methods along with their applications in industryThe remainder of the manuscript is organized as followsSection 2 gives an up-to-date overview of the methods forautomatic characterization of the visual appearance andproposes a taxonomy to group them in a meaningful wayParticular attention is given to robustness against noisefactors such as variations in illumination and viewpointSection 3 deals with the main applications in industry withparticular emphasis on grading surface inspection andcontent-based image retrieval (CBIR) Concluding remarksare reported in Section 4 along with a discussion on futureresearch directions and potentially innovative applications

2 Methods

Originally approaches to automatic characterization of thevisual appearance of materials considered colour and textureseparately In the last decade however an increasing interesthas been devoted to integrating these two sources of data intoa unifying model Strong evidence has in fact supported thisidea and confirmed that incorporating colour into the texturemodel has beneficial effects in many applications [17 18]Theclassification scheme that we propose in this paper reflectsthis idea namely that themethods can be based on one of thetwo features separately (either colour or texture) or considerthe two of them jointly

Based on these considerations itmakes sense to define thefollowing three main categories spectral spatial and hybridmethods [19] The first group is based on colour only therelative spatial distribution of the pixel values is not taken intoaccount The second group encloses texture-based methodsthey are called spatial methods since they take into accountthe relative variation of pixel intensity (grayscale values) inthe spatial domain but discards colour altogether (images arepreviously converted into grayscale) Finally those methodsthat consider colour and texture jointly are referred to ashybrid methods Before describing the three groups in detailwe will first discuss three points that apply to each method ingeneral these are dimensionality robustness to noise factorsand parameter tuning

Dimensionality refers to the number of parameters amethod relies upon Whichever the approach we use tocharacterize the visual appearance of a material this will berepresented through a finite set of quantitative parameterswhich are usually referred to as the feature vectorThe appear-ance of a material can therefore be regarded as a point in an119899-dimensional space where 119899 is the number of parametersIdeally a good method should provide high discriminationaccuracy with as few parameters as possible Computation infact becomes costly as dimensionality increases representinga serious limitation for applications that require real-timeprocessing Furthermore long feature vectors are likely todegrade the performance of a classifier as a consequence ofthe well-known phenomenon of ldquocurse of dimensionalityrdquo[20]

Noise factors represent uncontrollable sources of uncer-tainty that commonly occur in nonideal conditions Inreal working environments it can be quite complicated tomaintain steady and controlled imaging conditions Mostcommonly one has to deal with variable conditions such aschanges in rotation scale illumination and so forth It can bequite difficult to compensate for such sources of noise Someof them are likely to degrade the performance of a methodso strongly as to make it virtually useless To illustrate thisconcept Figure 1 shows how the appearance of a material canbe affected when exposed to different sources of noise

The effects of changes in illumination are reported inFigures 1(a) and 1(b) Changes in translation scale rotationandor viewpoint are shown in Figures 1(c) 1(d) 1(e) and1(f) As we discuss in the following subsections each classof methods is sensitive to one or more sources of noise Itis therefore the designerrsquos responsibility to select the methodwhich is most appropriate to the specific application domainconsidering the potential noise factors that can arise in theirspecific context

Finally it is important to mention that virtually anyspectral spatial or hybrid image descriptor depends onone or more parametersmdashsee Table 1 for a roundup of themethods presented here and their parameters In any specificapplication the performance of a method may vary greatlydepending on the setup of such parametersWhen it comes toimplementing industrial systems it is advisable to stick to thevalues suggested in the literature as a first step then to furtherimprove the results it is possible to optimise the parametersrsquovalues in a supervised fashion adaptive to thematerial to theimage kind but also to the experience of the user

In the following subsections we analyse spectral spatialand hybrid approaches to colour texture analysis and discusstheir pros and cons For each class we present a set ofrepresentativemethods and their use in three different indus-trial applications grading surface inspection and CBIR InTable 2 we summarise the robustness of each method todifferent noise factors (ie changes in illumination inten-sity viewpoint rotation and scale) and report a qualitativeappraisal of the computational time The methods includedin the experiments have been selected on the basis ofclassification accuracy computational efficiency and easinessof implementation


(a) (b) (c)

(d) (e) (f)




Spectral



Spatial



frequency ratio

Hybrid











Spectral

R G B means ∙



Colour centiles ∙

Spatial

GLCM ∙∙

LBPri81 ∙

CCRri81 ∙∙


Hybrid

OCLBPri81 ∙∙










(a) (b)






and CCRri81









81


81




81












Spectral



Spatial

GLCM 5 548 466 702 595LBPri

81 36 964 840 945 806CCRri

81 72 844 698 894 796Gabor filters 32 991 818 906 956

Hybrid

OCLBPri81 216 980 876 1000 962










1distance [20]


119886 =

1

119875

119875

sum

119901=1

119862119901

119881119901

(1)











Fabric

Paper

Wood

colour centiles

A = 9891 A = 9865 A = 9926

A = 9532 A = 9856 A = 9796

A = 9929 A = 9530 A = 9784


LBPri81



81



81

and LBPri81




119860 =

1003816100381610038161003816119861 cap 119861119879

1003816100381610038161003816

|119868|

+

1003816100381610038161003816119865 cap 119865119879

1003816100381610038161003816

|119868|

(2)


119879and 119865





(a) (b)








81


81


4 Conclusions







Spatial


LBPri81

LBPri81


Acknowledgments


References


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c)

(d) (e) (f)




Spectral



Spatial



frequency ratio

Hybrid











Spectral

R G B means ∙



Colour centiles ∙

Spatial

GLCM ∙∙

LBPri81 ∙

CCRri81 ∙∙


Hybrid

OCLBPri81 ∙∙










(a) (b)






and CCRri81









81


81




81












Spectral



Spatial

GLCM 5 548 466 702 595LBPri

81 36 964 840 945 806CCRri

81 72 844 698 894 796Gabor filters 32 991 818 906 956

Hybrid

OCLBPri81 216 980 876 1000 962










1distance [20]


119886 =

1

119875

119875

sum

119901=1

119862119901

119881119901

(1)











Fabric

Paper

Wood

colour centiles

A = 9891 A = 9865 A = 9926

A = 9532 A = 9856 A = 9796

A = 9929 A = 9530 A = 9784


LBPri81



81



81

and LBPri81




119860 =

1003816100381610038161003816119861 cap 119861119879

1003816100381610038161003816

|119868|

+

1003816100381610038161003816119865 cap 119865119879

1003816100381610038161003816

|119868|

(2)


119879and 119865





(a) (b)








81


81


4 Conclusions







Spatial


LBPri81

LBPri81


Acknowledgments


References


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation














Spectral

R G B means ∙



Colour centiles ∙

Spatial

GLCM ∙∙

LBPri81 ∙

CCRri81 ∙∙


Hybrid

OCLBPri81 ∙∙










(a) (b)






and CCRri81









81


81




81












Spectral



Spatial

GLCM 5 548 466 702 595LBPri

81 36 964 840 945 806CCRri

81 72 844 698 894 796Gabor filters 32 991 818 906 956

Hybrid

OCLBPri81 216 980 876 1000 962










1distance [20]


119886 =

1

119875

119875

sum

119901=1

119862119901

119881119901

(1)











Fabric

Paper

Wood

colour centiles

A = 9891 A = 9865 A = 9926

A = 9532 A = 9856 A = 9796

A = 9929 A = 9530 A = 9784


LBPri81



81



81

and LBPri81




119860 =

1003816100381610038161003816119861 cap 119861119879

1003816100381610038161003816

|119868|

+

1003816100381610038161003816119865 cap 119865119879

1003816100381610038161003816

|119868|

(2)


119879and 119865





(a) (b)








81


81


4 Conclusions







Spatial


LBPri81

LBPri81


Acknowledgments


References


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b)






and CCRri81









81


81




81












Spectral



Spatial

GLCM 5 548 466 702 595LBPri

81 36 964 840 945 806CCRri

81 72 844 698 894 796Gabor filters 32 991 818 906 956

Hybrid

OCLBPri81 216 980 876 1000 962










1distance [20]


119886 =

1

119875

119875

sum

119901=1

119862119901

119881119901

(1)











Fabric

Paper

Wood

colour centiles

A = 9891 A = 9865 A = 9926

A = 9532 A = 9856 A = 9796

A = 9929 A = 9530 A = 9784


LBPri81



81



81

and LBPri81




119860 =

1003816100381610038161003816119861 cap 119861119879

1003816100381610038161003816

|119868|

+

1003816100381610038161003816119865 cap 119865119879

1003816100381610038161003816

|119868|

(2)


119879and 119865





(a) (b)








81


81


4 Conclusions







Spatial


LBPri81

LBPri81


Acknowledgments


References


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation













81


81




81












Spectral



Spatial

GLCM 5 548 466 702 595LBPri

81 36 964 840 945 806CCRri

81 72 844 698 894 796Gabor filters 32 991 818 906 956

Hybrid

OCLBPri81 216 980 876 1000 962










1distance [20]


119886 =

1

119875

119875

sum

119901=1

119862119901

119881119901

(1)











Fabric

Paper

Wood

colour centiles

A = 9891 A = 9865 A = 9926

A = 9532 A = 9856 A = 9796

A = 9929 A = 9530 A = 9784


LBPri81



81



81

and LBPri81




119860 =

1003816100381610038161003816119861 cap 119861119879

1003816100381610038161003816

|119868|

+

1003816100381610038161003816119865 cap 119865119879

1003816100381610038161003816

|119868|

(2)


119879and 119865





(a) (b)








81


81


4 Conclusions







Spatial


LBPri81

LBPri81


Acknowledgments


References


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












Spectral



Spatial

GLCM 5 548 466 702 595LBPri

81 36 964 840 945 806CCRri

81 72 844 698 894 796Gabor filters 32 991 818 906 956

Hybrid

OCLBPri81 216 980 876 1000 962










1distance [20]


119886 =

1

119875

119875

sum

119901=1

119862119901

119881119901

(1)











Fabric

Paper

Wood

colour centiles

A = 9891 A = 9865 A = 9926

A = 9532 A = 9856 A = 9796

A = 9929 A = 9530 A = 9784


LBPri81



81



81

and LBPri81




119860 =

1003816100381610038161003816119861 cap 119861119879

1003816100381610038161003816

|119868|

+

1003816100381610038161003816119865 cap 119865119879

1003816100381610038161003816

|119868|

(2)


119879and 119865





(a) (b)








81


81


4 Conclusions







Spatial


LBPri81

LBPri81


Acknowledgments


References


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation














Fabric

Paper

Wood

colour centiles

A = 9891 A = 9865 A = 9926

A = 9532 A = 9856 A = 9796

A = 9929 A = 9530 A = 9784


LBPri81



81



81

and LBPri81




119860 =

1003816100381610038161003816119861 cap 119861119879

1003816100381610038161003816

|119868|

+

1003816100381610038161003816119865 cap 119865119879

1003816100381610038161003816

|119868|

(2)


119879and 119865





(a) (b)








81


81


4 Conclusions







Spatial


LBPri81

LBPri81


Acknowledgments


References


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b)








81


81


4 Conclusions







Spatial


LBPri81

LBPri81


Acknowledgments


References


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












Spatial


LBPri81

LBPri81


Acknowledgments


References


















































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









automatic characterization of the visual appearance of industrial

Documents