Automatic Machine Vision-based pilling measurement: a review Rocco Furferi1*, Lapo Governi1, Yary Volpe1 *Corresponding Author 1. Department of Industrial Engineering, University of Firenze (Italy). {
[email protected];
[email protected];
[email protected]}
Abstract Pilling is an undesired defect of textile fabrics, consisting of a surface characterized by a number of roughly spherical masses made of entangled fibers. Mainly caused by the abrasion of fabric surface occurring during washing and wearing of fabrics, this defect needs to be accurately controlled and measured by companies working in the textile industry. Pilling measurement is traditionally performed using manual procedures involving visual control of fabric surface by human experts. Since the early nineties, great efforts in developing automatic and non-intrusive methods for pilling measurement have been made all around the world with the final aim of overcoming traditional, visual-based and subjective, procedures. Machine Vision proved to be among the best options to perform such defect assessment since it provided increasingly performing measurement equipment and tools, serving the purpose of automatic control. In particular, a relevant number of interesting works have been proposed so far, sharing the idea of helping (or even replacing) traditional measurement methods using image processing-based ones. The present work provides a rational and chronological review of the most relevant methods for pilling measurement proposed so far. This work serves the purposes of 1) understanding whether today automatic machine vision-based pilling measurement techniques are ready for supplanting traditional pilling measurement and 2)
providing tthe textile technology t researcherss with a bird d’s eye view w about the main methods studiedd to confrontt with this problem. p Finally, onn the basis of the most relevant ssuggestions offered by y the review wed approacches, futuree trends in ppilling meassurement methods m are postulated so that the possible sccenario lying g ahead forr the future iis drawn. Keywordss: Fabrics, Pilling meaasurement, Machine Vision, V Imaage Processsing, Artificcial Neurall Networks.
Introdu uction As widelyy recognizedd [1], the term t “pillinng” is referrred to a su urface defecct occurring g in textilee fabrics andd consistingg of entang gled fibers forming th he so called d “pills”. Suuch pills are, usually,, caused by tthe combination of washing and w wearing of fabrics; f in detail, d due too the abrasio on of fabricc surface, a nnumber of loose l fibers tend to entaangle into short s fine haairs thus devveloping intto sphericall bundles annchored to thhe surface of o the fabricc (see Figuree 1).
Figure F 1–E Example of pilled p fabricc. The fabricc’s pills forrmation (i.e. the so callled “resistaance to pilliing”) is typpically meassured usingg proceduress described in Standard ds such as the D4970//D4970M-10e1 (ASTM M, 2010) an nd the UNII EN ISO 122945– 20044; since fabrrics take a loong time to o be pilled in n normal usse, resistancce to pillingg
needs to bee tested by a simulated d acceleratedd wear, follo owed by a visual v assesssment of th he degree off pilling baseed on a visuual compariison of the ssample to a set of test im mages. Two comm mon piecess of equipm ment for ppilling meassurement, mainly m useed in Europ pe, are thee Martindalee pilling tester and the Pilling Boxx. The Martinndale testerr consists of a numberr of testing plates (See Figure 2) oon which th he abradingg fabrics is attached; these t four testing t plattes are mou unted on th he base plaate of the instrument.. Generally speaking, fabrics f to bee tested usinng Martind dale are cut in an approoximate circular shapee t 90± 1 mm. m A worssted wool cloth is used d for abradiing the sam mples and a with diameeter equal to trajectory based on thhe Lissajou us figure iss used to perform each h cycle (m more preciseely, a cyclee consists off 16 movem ments in th he Lissajouus figure). A 12 kPa head presssure is appllied by thee machine.
Figure F 2–M Martindale pilling p testeer. In Pilling B Box (see Fiigure 3) sam mples are m mounted on polyurethan ne tubes andd are tumblled in cork-lined rotatiing woodenn boxes. Acccordingly, thhe samples move undeer the condittion of no pressure p andd the specim mens are conducted under u mutuual transien nt touching.. As a connsequence, unlike thee Martindalee method, thhe rubbing for f the sampples is random.
Figurre 3 – Pilling g Box. Whicheverr is the device (Martin ndale or Pillling Box), the t final ressult consistss of abraded d fabrics too be assesseed in terms of pilling. This is peerformed by y skilled op perators (exxperts) com mparing thee specimens,, after a prredefined nu umber of cyycles perfo ormed by th he testing eequipment, with visuall standards ((which mayy be actual fabrics f sam mples or pho otographs). On the basiis this comp parison, thee experts deffine the resistance to pilling p usingg the so callled “degreee of pilling”” i.e. an ind dex varyingg on a (arbittrary) scalee ranging frrom 5 -whhich means no pilling-- to 1 -whiich means very v severee pilling. This methood proves too be suitable for prediccting the acttual behavio or of fabric s everyday use only inn some speccific condittions. For instance, aaccording to t the ASM ME Standaard, laborattory test iss consideredd reliable as an indicatio on of relativve end-use performancce in cases w where the difference inn abrasion reesistance off various maaterials is laarge, but theey should not n be reliedd upon in prrediction off actual weaar-life in speecific end uses, unless tthere are daata showing g the specifiic relationsh hip betweenn laboratory abrasion tests and actu ual wear in tthe intended d end-use. According to [2], the main draw wback of thee subjectivee methods based b on esstimation by y experts iss their inconnsistency annd the inaccu uracy of thee rating resu ults. Hencefforward, theere is still to oday a needd for devisiing objective evaluattion methoods, relying g in autom matic and non-intrusive pillingg measuremeent.
With the aim of speeding up the pilling measurement procedure and, at the same time, to increase the reliability of the visual control, in the last years a number of Machine Vision (MV) systems have been proposed in order to overcome the limitations of traditional, visually-based, pilling measurement. On the basis of the most relevant results obtained in this field, the present paper provides a rational and chronological review of the most promising methods proposed so far. It is authors’ opinion that such a review can help researchers in understanding the working principles of today’s best automatic machine vision-based pilling measurement techniques. Moreover, on the basis of the best practices offered by the reviewed works, future trends in pilling measurement are postulated, so that interested researchers are aware of the future scenario that lies ahead for the future.
1. A categorization of methods for automatic fabric pilling measurement using Machine Vision In the last decades automated visual inspection (AVI) of fabrics for quality control faced an increasing trend in the textile industry due to the considerable development of technologies related to vision systems. Several approaches have been proposed in scientific literature [3-6] employing image processing-based methods and statistical parameters (such as mean, variance and median) for defect detection on fabrics. Pilling measurement using machine vision systems makes no exception: a number of methodologies have been proposed in order to explore automatic or semi-automatic pills detection and classification. Basically (in almost all the methods) the starting point consists of digital images of pilled fabrics. These images (representing either pilled fabric specimens to be evaluated or standard reference) are, then, processed in several different ways in order to extract some features describing fabric pilling. Finally, such parameters are used for grading the fabrics or for characterizing their quality. While the starting point and the final results are ultimately shared by all the techniques, what changes is the method adopted for extracting the information used for
pilling grading. On the basis of main literature works, in the present work the following categories are identified: 1) 2D imaging methods based on thresholding. 2) 2D imaging methods based on Fourier and/or Wavelet analysis. 3) 3D imaging methods. 4) AI-based methods (using either 2D or 3D images). Understandably, different categories could be used for describing existing works. Moreover some more recent techniques use approaches comprised in more than one of the above categories. Nevertheless, it is authors’ opinion that the given categorization, although open to improvement, is effective for understanding and systematizing the knowledge about how the pilling assessment problem has been faced by more than a few authors all over the world. As already mentioned, possible approaches are presented in a chronological order so that the main improvements brought by researchers are time-streamed. 1.1. 2D imaging methods based on thresholding The main idea of almost all the papers dealing with 2D imaging methods is to perform pills detection using image segmentation [1] i.e. the process of partitioning the original image into multiple segments including fabric background and pills. This process is usually, in its turn, aimed at determining parameters such as the number and the density of pills and/or the area occupied by the pills on the fabric surface. Once this task is performed, pilling grade is obtained as a parameter inferred from the number of pills, or by comparing the pilled fabric with a reference fabric (either with or without pills). As a matter of fact, almost all methods classified in this category use, at some point, an image binarization by applying one or more thresholds and, possibly, morphological operations on images. In Figure 4 an exemplificative flow diagram characterizing this category of methodologies is shown.
Figure 4 – Flow dia agram of 2D D imaging methods m bassed on thressholding. An early w work dealingg with imag ge segmentaation was caarried out by y Konda et aal. in 1990 [7]; imagess of fabric ssamples, pillled using Martindale M equipment,, were acqu uired using a commerccial cameraa under nearr-tangentiall illuminatio on thus obbtaining imaages with high h pill-too-backgroun nd contrast.. Obtained images are then t binarizzed using tw wo differentt thresholds with the fin inal result of o detachingg the backgrround from the pills. In Konda’s work, the background b d is represennted with black b (pixell value equaal to 0) whille pills are depicted d as white blobs (pixel valu ue equal to 1). Eventuaally, pillingg class of the fabric sam mple under investigatioon is evaluaated from the total num mber (or to otal area) off pills. In Figure 5 an illustrativee image froom Konda’ss work describing thee number of pills as a function off pill size is proposed.
Figure 5 – Numberr of pills as a function oof pill size: an examplee taken from m Konda’s work w [7]. In 1996, A Abril et al. [8] [ used som me techniqu ques typical of digital image i proceessing with h the aim off evaluating the pilling degree. Fro om the analyysis of a seet of standarrd images a sequential method forr an objectivve measurem ment was deevised. An iintermediatte result of the t proposeed approach h consists off binary imaage obtainedd using segm mentation bby local bin narization. In n Figure 6 a 64x64 pix xels portionn of such binnarized imagge, taken from Abril’s work, is shown.
Figure 6 – Part of a processed d image afteer segmenta ation by loca al binarizatiion (Abril et e al.[8]). Starting froom the binaary image, an a evaluatioon of the to otal pilled arrea (for eacch processed d image) too be related with the pilling p degrree is carrieed out. In particular, authors claaim that a logarithmicc t total pilling area annd the degreee of pilling g subsists obbserved (seee Figure 7).. relationshipp between the The proposed methodd has been further f impllemented by y the same authors in [9] by using g 1) a Top-hat transfoorm (an opeeration thatt extracts sm mall elemen nts and details from ggiven imagees [10]) forr obtaining bbackgroundd uniformity y, 2) an imaage segmenttation based d on a Gausssian model [11] of thee
background, and 3) a selective noise elim mination in the binary y image. Thhe maximu um error off misclassifiication in peercent of baackground ppixels in th he total amo ount of pixeels classified as pillingg (beyond thhe thresholdd) was found d equal to 0 .3% for an optimally o seelected threeshold valuee.
Figure 7 – Areas of pilling p corrresponding to the stand dard imagess of pilling ddegree varyying in the d image-proccessing bassed method range [1-55]: comparrison betweeen human viisual perforrmance and proposeed by April et e al. [8]. In [12] thee main conccept described in [8] w were recalled d and a digiital image pprocessing was w used too determine pills size, number, n shaape, orientattion angle, contrast, c tottal area andd the mean area a of pillss on a fabricc surface, esspecially usiing threshollding techniiques. A MV-bassed methodoology that automatical a lly counts the number of pills on textile fabrric sampless and classiffies them innto pre-defined classess has been proposed p in n [13]. A C CCD cameraa is used too capture suuccessive grray scale im mages of thhe fabric saample; then, segmentattion, Radon n transform m [14], morpphological filtering, f an nd de-trendiing operatio ons [15] aree applied too determine the pillingg
count. Usinng fuzzy meembership functions f [116], the fabrric pilling co ount is ultim mately relatted to fabricc pilling resiistance. A tool devveloped to detect and d describe ppills on so olid-shade fabrics f (afteer being im maged withh conventionnal personall-computer--based harddware) has been deviseed in 1998 [17]. In su uch a work,, the deviseed softwaree evaluated the total nnumber, to otal area, and total voolume of fabric f pills.. Moreover, the system m evaluated d distributioons of pill size, shapee, orientatioon angle, contrast andd uniformityy of pill spattial distributtion on the ffabric. In the same year, Xiaoohong and Mu M [18] prooposed a method m for pilling evaluuation wheree the imagee of pilled ffabric is preeprocessed on the basiis of imagee's gray-scalle statisticaal and/or mathematicall morphologgy. The pilling of fab bric is, thenn, assessed syntheticallly on the basis of th he size, thee number annd the morrphology off pilling. T Tested with h knitted saamples, the results pro oved to bee satisfactoryy. In 1999, Fazekas et all. [19] locatted pill regiions on fabrric samples by combinning templatte matchingg techniquess and imagge thresholding. Spe cial illumiination arraangements, i.e. a bi-directionall illuminatioon, were ussed to grasp p the depthh informatio on from images, so thhat pills were properlyy segmentedd from the background (see Figure 8).
Fiigure 8 – Bii-directiona al illuminatiion used forr pill-detectiion by Fazeekas et al [1 19] Finally, staatistically comparing th he number of pills dettected over a given areea with the assessmentt (quality cllassificationn) given by y the textilee experts, it is possib ble to emppirically dettermine thee
optimal thrreshold valuues - measu ured in pillls per area - between the t quality classes defi fined by thee standard. A remarkaable approacch to extracct pill featurres from fab bric images was propossed in [21 20]; 2 using a two-dimennsional Gauussian fit th heory, authoors train a “pill “ templaate'' using aactual pill images i andd determine a reasonablle threshold d for image ssegmentatio on using a histogram-fi h fitting techniique. Usingg the describbed approacch five paraameters to describe pill properties (i.e. pill nnumber, mean area off pills, total area of pills, contrast and densitty) are defin ned. Finally y, from succh data, a definition off pilling gradde is providded. The level of pilling has h been alsso identifieed and charaacterized ussing the sizze and num mbers of thee existing pillls in 2005 by b Huang et e al. [21]. Since segm mentation algorithms a can c be affeccted by fabric texture, color, and pattern, an n edge-flow w based algoorithm takinng all these factors intoo account haas been pro oposed in [222]. This ap pproach cann be used in different kinds k of fabrics, especiially those having h com mplex backgground. In Figure F 9 thee mentation obtained o in such a workk starting frrom fine tex xture wovenn is shown. pilling segm
n performed d in [21]. Figure 9 – Pilling s egmentation The final rresult of thee proposed method m connsists of properly segmented imagees where piills are easyy detectable from the background. b . Awkwardlly, no inforrmation reg garding the pilling grad de derivingg from imagee analysis iss provided.
A more reccent applicaation of imaage analysiss to assess the fabric wrinkle w andd abrasion reesistance inn order to coompare witth experimeental methoods is descrribed in [23 3]. By empploying an appropriatee lighting m method, sam mple imagess were capttured by ussing a scanner; then, iimages prepared from m samples w were processsed using MATLAB® M in order to o extract the pills from m the backg ground thuss deriving a pilling gradde. In [24] an edge-flow based b fabricc pilling seggmentation algorithm which w utilizees image co olor, texturee and phase of the edgee flow vecto or [25] was adopted in order to im mplement thhe pilling seegmentationn of various complex faabrics.
Figure 10 0 – Pilling ssegmentatio on performeed in [24]. After recoggnizing thee pilling fro om its backgground, thee total numb ber of pillinng can be obtained o byy searching tthe connectiive regions in binary im mage. As deepicted in Figure F 11, evvery connecctive regionn is labelled and the totaal number of o non-zero pixel values is calculatted.
matic repreesentation off the method d for search hing pillingg regions pro ovided in Figure 111 – A schem [24].
The relatioonships betw ween pilling g grades annd the total number off pilling, thee size of the total areaa and the opttical pillingg grading aree declared eequal to, resspectively 0.96, 0.94 annd 0.92. A novel m method for loocating the pills p in wovven fabric based on Gab bor filter [226] is proposed in [27];; fa imagees in order to t remove fabric f texturres, thus enh hancing thee Gabor filteer is appliedd to pilled fabric pills. In thee enhanced fabric imag ge, thresholdd method iss finally used to segmennt and locatte the fabricc pills. In [28] a m method to analyze a pilled p knittted fabric surface by using u colorr digital images (RGB B model) is pproposed. Application A B model forr the acquirred images ((see Figure 12) allowss of the RGB differentiatting pillingg from fuzzzing channges more effectively and preciisely with respect too grayscale-bbased methoods. The fin nal result off this approach, whose flow-chart is illustrateed in Figuree 13, consistts of an indeex N indicatting the perccentage of pilled p area. Such a valuue is lastly related r withh the fabric pilling graade. Moreov ver, the claassification of the pilling grade uusing the N value forr posed as a fu future work.. different grroups of fabbrics is prop
Figure 12 2 – Pilling ssegmentatio on performeed in [28]. The extendded mean shhift algorith hm was alsoo used to trry to solve the t segmenntation of faabric pillingg images in [29] by introducing g two mainn steps: im mage pre-fi filtering andd final seg gmentation.. formed by authors shhows that the proposed algorithhm can geet excellentt Laboratoryy test perfo segmentatiion if an opttimal choicee of the 3 reequired threeshold param meters is asssumed.
Figure 13 – Flow-charrt of the meethod proposed in [28].. 1.2. 2D im maging methods base ed on Fourrier analyssis and Wa avelet While the above menntioned papers are mosstly based on o image th hresholdingg, another raange of 2D D image proocessing baased methods developped for asssessing pillling grade is related to Fourierr Transform and Wavellet analysis (see Figuree 14).
Figuure 14 – Floowchart of 2D 2 imagingg methods ba ased on Fou urier analyssis and Wavvelet.
In 2002, Jeensen and Carstensen C [30] [ took ann image fro om fabric su urface and uused a Fourrier mask too filter the kknitted stitchh backgroun nd from thee fuzz and pill. p In partiicular, the FFourier massk has beenn used to filtter the knittted stitch background b from the fuzz fu and pill. A pillingg measurem ment cabinett was specifi fically desiggned and dev veloped in [[31] (see Figure 15).
Figurre 15 - pillinng measurem ment cabineet was specifically desiigned and ddeveloped in n [31]. Captured iimages werre analyzed d using apppositely dev veloped softtware basedd on thresh holding andd various pillling param meters such as total nuumber of pills, p total area a of thee pills, meaan area andd number off pills per unit u area aree measured.. Such paraameters werre, then, com mpared witth the samee ones obtainned manuallly, thus sho owing a goood correlatio on with fabric grading performed by experts.. In particulaar, authors demonstrated that the highest is th he pilling, the t larger iss the pills per unit areaa parameter. In Figure 16 some deetails aboutt the pilling g parameterrs of standaards obtaineed from thee system devvised in [311] and the EMPA E standdards are prroposed thu us demonstrrating the efffectivenesss of the propposed methood.
Figuree 16 - some details abou ut the pillinng parameteers of standa ards obtaineed from thee system devised in [31]] and the EM MPA standa ards.
A more recent approoach to pillling evaluaation based on the wavelet reconnstruction scheme wass investigateed in [32]. The metho od, preliminnary evaluaated using SM50 S Euroopean stand dard pillingg images, shhows that reconstruccted resoluution level, wavelet bases b and sub-imagee used forr reconstructtion affect the t segmen ntation of piills and, thu us, pilling grading. g Thhe area ratio o of pills too total imagee was succeessfully used d as a pillingg rating facctor (in analogy with a ggood number of workss belonging to the all thhe 3 categories mentionned before). main image processing is used to separate s perriodic structtures in the image (thee In [33] freqquency-dom fabric weavve/knit patttern) from non-periodic n c structures in the imag ge (the pills)). The authorrs propose that t for two o-dimensionnal discrete wavelet transform (2D DDWT) analysis of un-pilled fabriic images, where w the wavelet w scalle is close to t the fabricc inter-yarnn pitch, the distributionn of detail cooefficients will have a relatively small stand dard deviatiion. On thee other hand d, when thee amount of pilling incrreases, also the standardd deviation will increasse as the pillls introducee variationss mage that diisrupt the un nderlying ppattern of th he fabric stru ucture. Refferring, for instance, too into the im Figure 17, taken from m [33], it caan be noticeed that for fabrics with h pilling grrade equal to t 1 (in thee paper indiccated with the t letter i) a lower vallue for standard deviation (i.e. thee coefficient ܵܦܿܦ inn the paper) can be fouund. As the pilling graade rise from m 1 to 5 (in n the paperr from i to v), v also thee ܵܦܿܦ tendds to rise.
Figure 17 – Test image pill inteensity ratingg vs. standa ard deviation ܵܦܿܦ fo for five pillin ng grades froom i to v [33].
However, as stated by the autho ors, a drawbback of this method iss that frequuency domaain analysiss cannot proovide locattion inform mation. Mooreover, un nder particu ular condititions, pillin ng may bee expected too occur periiodically, so o that it cann nnot be easilly discrimin nated. A new appproach for pilling p evalu uation basedd on the mu ulti-scale tw wo-dimensioonal dual tree complexx wavelet traansform (CW WT) has beeen proposeed in [34]. The T CWT method m [35]] is used to decomposee the pilled fabric imagge with six orientationns at differeent scales an nd reconstrruct fabric background b d texture andd pilling suub-images. An A energy analysis method m is, att that time, used to search for ann optimum iimage decoomposition scale and tto dynamiccally discrim minate pillin ing image from f noise,, fabric textuure, fabric surface s unev venness andd brightnesss variation in n the pilled fabric imag ge. In Figure 118 a 3D meesh plot of WoolMark® W ® SM50 Grrade 1 woveen fabric is shown, tak ken by [34].. Using the pproposed method m it is possible to identify pilling inform mation over a fused and d smoothedd background of gray value v zero att different sscales. The positive an nd negative maximum gray g valuess d imagee represent tthe highest point of pilling and thhe deepest point p of thee of the recoonstructed detail pilling shaddow respectively.
Figure 188 – 1) 3D mesh m plot off WoolMark® ® SM50 Grrade 1 woveen fabric froom [34]; 2) identified pillingg; 3) identifi fied pilling aat scale 5; 4) 4 identified d pilling at sscale 6.
This approach can be considered hybrid with the ones described in Section 1.4. since a LevenbergMarquardt back-propagation neural rule is finally used to classify the pilling grade. The robustness of the above proposed method has been assessed by Zhang et al. in 2012 [36]. In detail robustness in terms of image rotation, image dilation, image brightness variation and image contrast variations has been assessed. The results provided by the authors suggest that the pilling identification method is robust to significant variation in the brightness and contrast of the image, rotation of the image and dilation of the image. The pilling feature vector developed to characterize the pilling intensity is robust to the brightness change (but sensitive to large rotations of the image). Obviously, it requires all images be arranged such that the illumination is coming from the same direction. As long as all images are adjusted to have the same contrast level, the method provides an objective measurement of the pilling volume and so it can be used to classify the pilling intensity. 1.3. 3D imaging methods To comprehensively understand fabric pilling phenomena thus allowing an accurate pilling grading, 3-D non-contact scanning systems can be considered better equipment with respect to 2D imaging devices. Actually, these devices are able to evaluate the overall fabric surface unevenness as well as pill characteristics (e.g. pill number, area, and population density). On the basis of this awareness, some recent works have been carried out in order to assess pilling grading of fabrics (see Figure 19) using 3D vision. In fact, at the beginning of the 21th century, Sirikasemlert and Tao [37] described an objective evaluation system for characterizing textural changes in knitted fabrics during simulated wear using 3D devices and, in particular, laser-triangulation technique. The investigated textural changes are pilling, fuzzing, and changes in the constituent yams. Two-dimensional Fourier analysis and wavelet analysis are, then, introduced as new analytical tools for objectively evaluating surface textural changes based on 3D features.
Figuree 19 – Flowcchart of 3D D imaging methods. m Another im mportant atttempt in usiing 3D imagging for pillling assessm ment is proovided in [38], where a CCD cameera was agaain used to capture c the iimage of a laser line projected onn the surfacee of a seriess of fabric sppecimens. Then, T makin ng use of trrigonometriic calculatio ons, the threee-dimensio onal shapess of the insppected fabriccs are recov vered. Suchh a 3D recon nstruction is eventuallyy used for determining d g the numberr, the area, and a the den nsity of pillss. The studyy proposed by Kim and a Park [339] is, perrhaps, the first f attemppt in quanttifying andd evaluating fabric pillling using g alternativvely two-dimensional and three--dimensionaal imagingg methods deepending onn the kind of o base fabriic to be exam mined. Two o-dimensionnal imaging g equipmentt consisting of a CCD camera c (640 0x480 pixelss) is used fo or capturing g a square arrea of 50 mm m for somee samples. 22D acquisittion is then n integratedd with a th hree-dimenssional meassurement obtained o byy means of thhe equipmeent shown in n Figure 20.. Various im mage processsing techniques such as FFT-based ones, as a well as tthree-dimen nsional dataa processingg algorithmss, were dev veloped for extracting both pills informationn and a seriees of shapee parameterss, ultimatelyy used for the t objectivve evaluatio on of fabric pilling. Auuthors finallly state thatt fabrics witth thick andd hairy yarn n are better assessed ussing three-d dimensional measuremeent becausee the sizes off pills are raather large.
F Figure 20 – schematic diagram d of the 3D mea asurement syystem propoosed by [39 9]. On the othher hand, faabric with th hin and sm mooth yarn seem s unsuittable for thiis method because b thee size of pillls sometimees went belo ow the intriinsic measu urement erro or of the 3D D system an nd thereforee it is necesssary to use only o 2D-bassed techniquues. Mettes et al. [40] deevised a method m baseed on opticcal triangullation that performs topographic t c reconstructtion of textile fabric saamples. Thee approach allows the evaluation and the qu uantificationn of the pilling formatioon on the baasis of topoographic chaanges. Seveeral tests hav ave been carrried out byy authors prooviding robuust and preccise results. Xu et al. [[41] introduuced a 3D im maging sysstem design ned for objeective evaluuation of fab bric pilling.. The system m was aimeed at reconsstructing 3D D surfaces of o fabrics by b using tw wo side-by-sside imagess captured byy a pair of digital d cameeras, withouut special lighting. Oncce the depthh data are caaptured, thee most relevvant inform mation for pilling seggmentation is identified, and piilling is su ubsequentlyy measured. A non-conntact methodd based on laser trianggulation sen nsor capablle of measuuring the heeight of thee fabrics witth five microon accuracy y has been ddeveloped by b Saharkhizz and Abdoorazaghi in 2012 2 [42]. For each ssample, 625500 data points were ccollected an nd in order to remove nnoise invollved, data a median alggorithm for smoothnesss is applied (see Figuree 21).
Figure 211 – Examplee of raw datta acquired ffrom a fabrric using thee device desscribed in [42] [ vs. the sam me data filteered using a median fillter. Once acquuired, a set of pilled faabrics was processed using u Median-cut, K-m mean and Competitive C e Learning aalgorithms in order to extract e the nnumber of pills, p protrud ding geomeetrical volum me and areaa of pills. Thhe comparison among the three aabove menttioned meth hods revealeed that the correlationn factor in oobjective annd subjectiv ve evaluatioon of the pilling p of th he samples by using Competitive C e Learning oof K-mean algorithm is 0.985 aand is morre reliable when comppared to K-mean K andd Median-cuut methods (see ( Figure 22).
Figure 22 - comparisson among correlation c ffactor of ob bjective and d subjective ffabric pill grading g by ussing Mediann-cut, K-meean, Compettitive Learn ning of K-meean methods ds (from [42] 2]). 3D reconstruction of fabric surffaces (incluuding the piills) has alsso been reccently realizzed in [43]] using gradiient fields method m and starting froom a set of 4 images. Afterward, A th the pills are detected inn 3D fabric ssurfaces usiing image-p processing toools, typicaally implemeented in thee MATLAB B® softwaree environmeent.
Recently, [[44] a stereoovision systtem and thee three-dimeensional (3--D) image aanalysis alg gorithms forr fabric pillinng measureement has also been prooposed (seee Figure 23)). Based onn the depth information i n available inn the 3-D im mage, the piilling detecttion processs starts from m the seed seearching at local depthh maxima to the region growing aro ound the seelected seeds using both h depth andd distance crriteria.
Figure 223 – a) Illusstration of the t stereovission system m proposed in i [44]; b) ggenerated 3D 3 fabric image. After the ppilling deteection, the density, he ight, and area a of individual pillss in the im mage can bee extracted too describe the t pilling appearance, a as shown in n Figure 24.
Figure 24 4 – Map of ddetected pillling areas from f [44]. 1.4. Artifiicial Neura al Network k based ap pproaches Artificial N Neural Netw works (ANN Ns) has beenn extensivelly used at a research staage in severral fields off textile induustry since last decad des [45-61].. ANNs aree particularly applied in spinning g, weaving,, dyeing andd quality coontrol engin neering. Thhe massive use of thiss frameworkk encouraged also thee developmeent of methoods for pilliing evaluatiion. In detaail, ANNs are a combineed together with imagee
processingg in order too strengthen n the traditioonal approacches. The main m idea, fo for this rang ge of works,, may be twoofold: 1) ANNs m may be usedd as the finaal phase afteer the imagee segmentattion (2D or 3D) for corrrelating thee pills typicaal parameterrs with the experts’ e juddgment (seee Figure 25).
Figure 255 – Flowchaart of ANNss-based metthods wheree ANNs are used in thee final stage of pilling grading. 2) ANNs m may be usedd to avoid im mage segmeentation forr detecting the t pills; fabbric imagess are treatedd using imagge processinng methodss with the aaim of extraacting significant param meters to be used as a training sett for intelliggent system ms (but not tto segment image); con nsequently, if the startiing phase iss still imagee processingg, this is used u not foor separatin ng pills from m the backkground bu ut as a pre-processingg step for AII-based metthods (see F Figure 26).
Figure 266 – Flowchaart of ANNss-based metthods wheree ANNs are used for coorrelating pa arameters (not neceessarily relaated to pills or coming ffrom imagee segmentattion) with thhe experts’ judgment. j Dealing wiith the first kind of meethods, in [662] Chen an nd Huang ev valuated annd graded faabric pillingg based on light projecction using g image annalysis and neural network to ovvercome th he commonn difficulty oof interferennce with fab bric pill infformation frrom fabric color c and ppattern. Pilliing grade iss assessed byy using a Kohonen K sellf-organizinng feature map m [5] neural networkk. Thirty diffferent kindd of pilled ffabric sampples were trained andd tested, an nd the corrrelation cooefficient between thee objective ggrade and suubjective grrade is 0.944 for the traaining samp ples and 1 fo for the testin ng samples.. The samplle number with w ± 1 grrade deviatiion is 5, so o the objecttive inspecttion accuraccy is statedd equal to 833%. In [63] ann energy annalysis metthod to seaarch for an n optimum image deccomposition n scale andd dynamicallly discriminnate pilling g image from m noise, faabric texturee, fabric sur urface uneveenness, andd illuminativve variationn in the pilled fabric image is proposed. For pillingg objective rating, sixx parameterss were exttracted from m the proccessed imag ges to desscribe pill properties. Finally, a Levenbergg–Marquarddt back prop pagation neuural rule was w used as a classifierr to classify y the pillingg
grade. Thee proposed method was w evaluateed using kn nitted, wov ven, and noonwoven pilled fabricc images acqquired by a digital camera. In Xin’s PhhD dissertattion [64] the results obbtained in [2 21] were useed as a trainning set for three kindss of ANNs iin order to correlate c the extracted parameterss with the expert’s judggment. Resu ults show a correlationn coefficientt equal to 0..9358. Dealing w with to the second kind d of methoods, in [65]] a feed forrward backk-propagatio on artificiall neural netw work (FFBP P ANN) hass been traineed to determ mine the degree of pillling (i.e. to classify thee fabric intoo a number of quality classes). IImages of pilled p fabricc samples aare acquireed using ann appositely devised maachine visio on system. T Then, the accquired imaages are pree-processed d in order too 1) discard color infoormation, 2)) to correctt non-unifo ormity in illlumination and 3) to extract 11 parameterss describingg pilled fabrics (see Figgure 27).
Figgure 27 –Fllow diagram m of the pro ocedure prop oposed in [665].
As mentioned above, such parameters are used for training the ANN which is finally used as a tool for assessing the degree of pilling of new fabric samples. The work is inspired by a previous one by the same authors [5] whose aim was to detect and classify a number of defects possibly occurring on raw fabrics such as stains, thin and thick bars, fillings, double fillings, weft threads, double warp threads and broken ends.
Conclusions From the state of the art analysis, it is evident that 2D imaging systems for pilling measurement are still used nowadays focusing on different, and progressively more precise and reliable, methods aimed at image segmentation. Whichever the image processing-based method for enhancing the pills is, almost all the methods share the idea that the central issue for automatic and objective pilling assessment is to directly measure pills geometric properties. This consideration can be, roughly, confirmed by analyzing the works summarized in Sections 1.1 and 1.2. Nonetheless, as described in Sections 1.3 and 1.4, 2D imaging is not the only possible strategy: in the last years a number of works implementing 3D imaging and AI-based systems have been devised in support or in substitution to “traditional” 2D-based methods. In particular, 3D imaging has been, explored, and still is, for developing pilling measurement methods taking into account the third dimension in the analysis. Since fabric pills have a 3D structure, such methods may be considered a promising frontier for textile control, especially thanks to the decreasing cost of 3D imaging devices. On the other hand, if on one side ANN-based methods are not new in the textile field, their application in pilling measurement (especially regarding pilling grading) is far to be fully explored and relevant works were carried out only in the last 10 years. As a consequence, in authors’ opinion, it is highly advisable to investigate more thoroughly methods which make use of these techniques. Moreover, it is necessary to favor more integration between imaging systems and AI-based ones.
In Table 1, the most relevant works are synoptically proposed divided into 4 periods: 1990 – 2000, 2001-2005, 2006-2010 and 2011-2014. From this time analysis, the following considerations can be drawn: Table 1 - most important works dealing with pilling measurement.
2D imaging Methods based on thresholding
1990-2000
2001-2005
Konda et al. 1990
Jensen et al., 2002
Abril et al. 1996
Xin et al., 2002
Annis, 1996
Behera et al., 2005
2006-2010
2011-2014 Gao et al., 2011
Naderpour et al., 2009
Abril et al. 1997 Jing et al., 2011
Xiaohong and Mu, 1998 Huang et al., 2005
Xiaojun et al., 2009
His et al. 1998 Fazekas et al., 1999
2D imaging Methods based on Fourier analysis and wavelet
Xin et al., 2002
Behera et al., 2005 Sirikasemlert and Tao, 2000 3D imaging Methods
Palmer et al., 2009
Jasińska, 2009
Kim and Park, 2006
Zhang et al., 2012
Xu et al., 2011 Saharkhiz and Abdorazaghi, 2012
Kang et al., 2004 Kim and Kang, 2005
Deng et al., 2011
Mendes et al., 2010
Mettes et al., 2006
Techniková et al., 2013 Behera and Mishra, 2006
AI-Based methods
Chen and Huang, 2004
Deng et al, 2009 Xin, 2009 Furferi et al., 2014
- after a strong initial impetus in the years 1990-2000, 2D imaging methods have not experienced a shutdown, and presumably more methods will be devised in the near future; - Fourier and wavelet analysis will probably be further investigated for pilling grading systems, since they are relatively new techniques applied to this field;
- 3D methods will probably replace some of the less promising 2D imaging methods in the near future; - ANN based methods were developed fairly late, as a consequence future developments are expected in this field. Summing up, the present work presented an overview of the most important works in the field of pilling measurement techniques and methods using MV. Without claiming to be exhaustive, this work is a first attempt in providing a survey of possible approaches dealing with this well-known field of textile research. Since one of the main problems preventing a direct comparison of the performance of the (many) existing methodologies is the lack of tests performed on the same set of specimens, future work will deal with direct implementation of the most promising techniques and their application on a broad set of fabric samples.
References [1] J- Fan, J., W.W. Yu, L. Hunte (2004). Clothing appearance and fit: science and technology. Cambridge: Woodhead Publishing Series in Textiles No. 33. pp. 54–60. [2] P. Gunavathi, K. Ragunathan (2008). Pilling evaluation: A new method. The Indian Textile Journal August 2008 issue. [3] Abouelela, A., Abbas, H. M., Eldeeb, H.,Wahdan, A. A., Nassar, S. M. (2005). Automated vision system for localizing structural defects in textile fabrics. Pattern Recognition Letters, Vol. 26(10), pp.1435-1443. [4] Carfagni, M., Furferi, R., Governi, L. (2005). A real-time machine-vision system for monitoring the textile raising process. Computers in Industry, 56 (8-9), pp. 831-842. [5] R. Furferi, L. Governi (2008). Machine vision tool for real-time detection of defects on textile raw fabrics. Journal of the Textile Institute, 99 (1), pp. 57-66.
[6] R. Furferi, L. Governi, Y. Volpe (2012) A novel method for ring spinning performance evaluation based on Computer Aided analysis of yarn geometry. International Journal of Mechanics, 6 (4), pp. 212-221. [7] Konda, A., Xin, L. C., Takadera, M., Okoshi, Y., Toriumi, K. (1990). Evaluation of pilling by computer image analysis. Journal of the textile Machinery Society of Japan, Vol. 36(3), pp. 96-107. [8] H.C., Abril, M.S. Millan, R.B. Navarro (1996). Pilling evaluation in fabrics by digital image processing. Proceedings of SPIE - The International Society for Optical Engineering, 2786, pp. 1928. [9] H.C. Abril, M.S. Millan, Y.M. Torres, R.B. Navarro (1997). Image segmentation based on a Gaussian model applied to pilling evaluation in fabrics. Proceedings of SPIE - The International Society for Optical Engineering, 3101, pp. 283-291. [10] R.C. Gonzalez, R.E. Woods, Digital Image Processing, Addison-Wesley, Reading (Mass) 1992. ISBN 0-201-50803-6 [11] J. Portilla, V. Strela, M.J. Wainwright and E.P. Simoncelli (2003). Image denoising using scale mixtures of Gaussians in the wavelet domain. Image Processing, IEEE Transactions on, 12(11), 1338-1351. [12] P. A. Annis, (1996). Pilling Evaluation of Laboratory Abraded, Laundered, and Worn Fabrics Using Image Analysis. Book of papers International Conference and Exhibition American Association of Textile Chemists and Colorists, pp. 465-479. [13] Iqbal M. Dar, W. Mahmood, G. Vachtsevanos (1997). Automated pilling detection and fuzzy classification of textile fabrics. Proc. SPIE 3029, Machine Vision Applications in Industrial Inspection V, 26 (April 15, 1997).
[14] L. Donald, The Radon transform on Euclidean space, Communications on Pure and Applied Mathematics 19.1 (1966), pp. 49-81. [15] D.S.G. Pollock, Trend estimation and de-trending via rational square-wave filters, Journal of Econometrics 99.2 (2000), pp. 317-334. [16] M.S. Chen and S.W. Wang, Fuzzy clustering analysis for optimizing fuzzy membership functions. Fuzzy Sets and Systems, (1999), 103(2), pp. 239-254. [17] C. H. Hsi, R. R. Bresee, P. A. Annis (1998). Characterizing Fabric Pilling by Using Imageanalysis Techniques. Part I: Pill Detection and Description. Journal of The Textile Institute Vol. 89, Iss. 1. [18] Wang, Xiaohong, Yao, Mu (1998). Assessing the pilling of fabric by image analysis- Journal of Dong Hua University (English Edition), 15 (4), pp. 68-70. [19] Z. Fazekas, J. Komuves, I. Renyi, L. Surjan (1999). Towards objective visual assessment of fabric features. Image Processing and its Applications, Vol. 1, pp. 411-416. [20] K. L. Jensen, J.M. Carstensen (2002). Fuzz and pills evaluated on knitted textiles by image analysis. Textile Research Journal, Vol. 72, No. 1, 2002, p. 34-38. 4, 2002. [21] L. Huang, G. Hong, R. Luo (2005). Fabric pilling assessment by digital image processing. 2005 Beijing International Conference on Imaging: Technology and Applications for the 21st Century, 2005, pp. 168-169. [22] Z.T. Xiao, H.W. Yang (2007). Fabric Pilling Segmentation Based On Edgeflow Algorithm. Proceedings of the Sixth International Conference on Machine Learning and Cybernetics, Hong Kong, 19-22 August 2007.
[23] F. Naderpour, S.A. Mirjalili, and M. Sharzehee (2009). The Investigation on the influence of DMDHEU on the Wrinkle and Abrasion Resistance of Cotton Fabrics using Image Processing Textile Research Journal November 2009 79: 1571-1577. [24] L. Xiaojun, H. Huabing, L. Yushu, Z. Hong (2009). The evaluation system of fabric pilling based on image processing technique. In International Conference on Image Analysis and Signal Processing, 2009. IASP 2009. (pp. 44-47). IEEE. [25] W.Y., Ma and B.S. Manjunath, Edge flow: a framework of boundary detection and image segmentation, In Computer Vision and Pattern Recognition, 1997. Proceedings., 1997 IEEE Computer Society Conference on (pp. 744-749). IEEE. [26] F. Bianconi, A. Fernández (2007). Evaluation of the effects of Gabor filter parameters on texture classification. Pattern Recognition, 40 (12), pp. 3325-3335. [27] W. Gao, S., Wang, R. Pan, J. Liu (2011). Automatic location of pills in woven fabric based on Gabor filter. Key Engineering Materials, 464, pp. 745-748. [28] J. Izabela (2009). Assessment of a Fabric Surface after the Pilling Process Based on Image Analysis, FIBRES & TEXTILES in Eastern Europe 2009, Vol. 17, No. 2 (73) pp. 55-58. [29] J. Jing, X. Kang (2011). Fabric pilling image segmentation based on mean shift. Communications in Computer and Information Science, 143 CCIS (PART 1), pp. 80-84. [30] B. Xin, J. Hu, H. Yan (2002). Objective evaluation of fabric pilling using image analysis techniques. Textile Research Journal, Vol. 72(12), pp. 1057-1064. [31] B.K. Behera, T.E., Madan Mohan (2005). Objective measurement of pilling by image processing technique. International Journal of Clothing Science and Technology, 17 (5), pp. 279291.
[32] S.C. Kim, T.J. Kang (2011). Image Analysis of Standard Pilling Photographs Using Wavelet Reconstruction Textile Research Journal December 2005 75: 801-811. [33] S. Palmer, X. Wang (2012). Classification of fabric pilling by image analysis. In EMAC 2003 proceedings: proceedings of the Sixth Engineering Mathematics and Applications Conference: University of Technology, Sydney, Australia, 9-11 July 2003 (pp. 175-180). Engineering Mathematics Group. [34] Z. Deng, L. Wang, X. Wang (2011). An integrated method of feature extraction and objective evaluation of fabric pilling. Journal of the Textile Institute, 102 (1), pp. 1-13. [35] I.W. Selesnick, R.G. Baraniuk and N.C. Kingsbury, The dual-tree complex wavelet transform. Signal Processing Magazine, IEEE, (2005), 22(6), pp. 123-151. [36] J. Zhang, X. Wang, S. Palmer (2012). The robustness of objective fabric pilling evaluation method. Fibers and Polymers, 10 (1), pp. 108-115.. [37] A. Sirikasemlert, X. Tao (2000). Objective evaluation of textural changes in knitted fabrics by laser triangulation. Textile Research Journal, 70(12), 1076-1087. [38] T.J. Kang, D-H. Cho, S.M. Kim (2004). Objective Evaluation of Fabric Pilling Using Stereovision, Textile Research Journal, Vol. 74(11), pp.1013-1017. [39] S. Kim, C.K. Park (2006). Evaluation of fabric pilling using hybrid imaging methods. Fibers and Polymers, 7 (1), pp. 57-61. [40] A.O. Mettes, P.T. Fiadeiro, R.A.L. Miguel (2006). Three-dimensional surface reconstruction for evaluation of the abrasion effects on textile fabrics. Proceedings of SPIE - The International Society for Optical Engineering, 6056, art. no. 60560F. [41] B. Xu, W. Yu, R. Wu Wang (2011). Stereovision for three-dimensional measurements of fabric pilling. Textile Research Journal December 2011 vol. 81 no. 20 2168-2179.
[42] S. Saharkhiz and M. Abdorazaghi, The Performance of Different Clustering Methods in the Objective Assessment of Fabric Pilling, Journal of Engineered Fibers and Fabrics , 7(4), (2012), pp. 35-41. [43] L. Techniková, M. Tunák, J. Janáček (2013). Determination and Comparison of Fabric Pills Distribution Using Image Processing and Spatial Data Analysis Tools, World Academy of Science, Engineering and Technology International Journal of Information Science and Engineering Vol:7 No:10, 2013. [44] Wenbin Ouyang, Rongwu Wang, Bugao Xu, Fabric pilling measurement using threedimensional image Journal of Electronic Imaging 22(4), 043031 (Oct–Dec 2013). [45] L. Di Angelo, P. Di Stefano (2014). An evolutionary geometric primitive for automatic design synthesis of functional shapes: The case of airfoils. Advances in Engineering Software, 67. [46] R. Furferi, L. Governi, Y. Volpe (2012). Modelling and simulation of an innovative fabric coating process using artificial neural networks. Textile Research Journal, 82 (12), pp. 1282-1294. [47] M. Dong, H.Y. Jiang (2010). Fabric defects image filtering method based on rough set and PC neural networks. Advanced Materials Research, 121-122, pp. 1012-1017. [48] A. Oonsivilai, N. Meeboon (2009). Silk texture defect recognition system using computer vision and artificial neural networks. Proceedings of the 2009 2nd International Congress on Image and Signal Processing, CISP'09, art. no. 5303972 [49] R. Furferi, L. Governi, Y. Volpe (2011). Neural network based classification of car seat fabrics. International Journal of Mathematical Models and Methods in Applied Sciences, 5 (3), pp. 696-703.
[50] P. Malik, D. Sarkar, B. Bagchi, S.S. Saha, A. Majumdar (2003). Application of artificial neural network for the modelling of yarn properties - A review. Man-Made Textiles in India, 46 (2), pp. 59-64. [51] R. Shamey, T. Hussain (2003). Artificial intelligence in the colour and textile industry. Review of Progress in Coloration and Related Topics, 33, pp. 33-45. [52] A. Kumar (2003). Neural network based detection of local textile defects. Pattern Recognition, 36 (7), pp. 1645-1659. [53] R. Stojanovic, P. Mitropulos, C. Koulamas, Y. Karayiannis, S. Koubias, G. Papadopoulos (2001). Real-time vision-based system for textile fabric inspection. Real-Time Imaging, 7 (6), pp. 507-518. [54] K.C. Fan, Y.K. Wang, B.L. Chang, T.P. Wang, C.H. Jou, I.F. Kao (1998). Fabric classification based on recognition using a neural network and dimensionality reduction. Textile Research Journal, 68 (3), pp. 179-185. [55] H. Liu, B. Liu, C. Feng (2011). Predictive modeling and empirical analyses of regional silk fabrics price index based on BP neural network. Advanced Materials Research, 331, pp. 685-688. [56] A. Majumdar (2011). Modelling of thermal conductivity of knitted fabrics made of cottonbamboo yarns using artificial neural network. Journal of the Textile Institute, 102 (9), pp. 752-762. [57] V. Mozafary, P. Payvandy, S.M. Bidoki, R. Bagherzadeh (2013). Predicting the influence of seam design on formability and strength of nonwoven structures using artificial neural network. Fibers and Polymers, 14 (9), pp. 1535-1540. [58] J.M. Rosa, K.R.M. Prado, W.A.L. Alves, F.H. Pereira, J.C.C. Santana, E.B. Tambourgi (2013). Applying of a neural network in effluent treatment simulation as an environmental solution for textile industry. Chemical Engineering Transactions, 32, pp. 73-78.
[59] S.D. Bhambure, A.J. Dhavale, P.V. Kadole, D.V. Kodavade (2013). Artificial neural network & its applications in textiles. Journal of the Textile Association, 74 (1), pp. 31-37. [60] S. Falomi, M. Malvezzi, E. Meli, A. Rindi. Determination of wheel
rail contact points:
comparison between classical and neural network based procedures Meccanica, 44, pp. 661
686,
2009. [61] B. Allotta, E. Meli, A. Ridolfi, A. Rindi. Development of an innovative wheel rail contact model for the analysis of degraded adhesion in railway systems. Tribology International, 69, pp. 128
140, 2013.
[62] X. Chen, X.B. Huang (2004). Evaluating Fabric Pilling with Light-Projected Image Analysis. Textile Research Journal, Vol.74, No.11, pp. 977-981. [63] Z. Deng, L. Wang, X. Wang (2011). An integrated method of feature extraction and objective evaluation of fabric pilling. Journal of The Textile Institute Vol. 102, Iss. 1, 2011. [64] B. Xin (2009). Characterization of fabric appearance based on image analysis, Institute of Textiles & Clothing, The Hong Kong Polytechnic University, Hong Long, 2009. [65] R. Furferi, L. Governi, Y. Volpe. (2014, in press) Towards automated and objective assessment of fabric pilling. International Journal of Advanced Robotics Systems.