research article fabric defect detection using local...

Research ArticleFabric Defect Detection Using Local HomogeneityAnalysis and Neural Network

Ali Rebhi Issam Benmhammed Sabeur Abid and Farhat Fnaiech

Scientific Research Laboratory in Signal Image Processing and Energy Mastery (SIME) University of Tunis5 Avenue Taha Hussein 1008 Tunis Tunisia

Correspondence should be addressed to Ali Rebhi ali rebhiyahoofr

Received 12 August 2014 Accepted 8 December 2014

Academic Editor Ivan Moreno

Copyright copy 2015 Ali Rebhi et alThis is an open access article distributed under the Creative CommonsAttribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

In the textile manufacturing industry fabric defect detection becomes a necessary and essential step in quality control Theinvestment in this field is more than economical when reduction in labor cost and associated benefits are considered Moreover thedevelopment of a wholly automated inspection system requires efficient and robust algorithms To overcome this problem in thispaper we present a new fabric defect detection scheme which uses the local homogeneity and neural network Its first step consistsin computing a new homogeneity image denoted as119867-image The second step is devoted to the application of the discrete cosinetransform (DCT) to the 119867-image and the extraction of different representative energy features of each DCT block These energyfeatures are used by the back-propagation neural network to judge the existence of fabric defect Simulations on different fabricimages and different defect aspects show that the proposed method achieves an average accuracy of 9735

1 Introduction

Defect detection is one of the main steps in quality control ofmanufacturing processes In the fabric field defect detectionbecomes an important task due to the widely used materialin daily life Fabric defects are responsible for nearly 85 ofthe defects found in the garment industry [1] Also it hasbeen observed [2] that price of textile fabric is reduced by45 to 65 due to defects It is imperative therefore todetect identify andprevent these defects from reoccurringAfabric defect is generally represented by awired variationwithrespect to the overall surface aspect Visual human inspectionleads to usual errors due to several reasons namely thefatigue and fine defects Hence the automation of theseoperations helps to improve quality and reduce labor costsAs confirmed by [3 4] there are more than 70 kinds offabric defects defined by the textile industry A human fabricinspection achieves a success rate about 60 to 75 [3] Figure 1shows the same machine automated inspection

Automated visual inspection methods are increasinglytaken into account in recent years There are various visualinspection systems such as defect detection of tiles [5] woods[6 7] ceramic [8] sheet steel [9] and textiles [10ndash19]

Fabric defect detection is one of the most schemingproblems in visual inspection Currently there are a lot ofnew studies and researches in fabric inspection [10ndash19] Thispaper emphasizes this problem and investigates some newtechnique to solve this problem

11 Related Work Texture analysis has a great interest inimage processing and there are many applications in itsanalyzing and processing one of which is defect detectionTherefore numerous methods are typically developed tosolve the defect detection task in fabric images The authorsin [4] propose a review of the art field in this state Chan andPang [10] provided a comprehensive overview of fabric defectdetection by Fourier analysis Wavelet transform is anothermethod which was applied in [11ndash14]

On the other hand Latif-Amet et al [15] proposed anapproach that uses wavelet theory and cooccurrence matrixfor the detection of defects encountered in textile imagesand classifies each subwindow as defective or nondefectiveby a Mahalanobis distance Gabor filter has been widelyused in the field of fabric defect detection All the detectionapproaches using Gabor filter can be classified into twocategories One is to use a filter bank such as [16 17] and the

Hindawi Publishing CorporationJournal of PhotonicsVolume 2015 Article ID 376163 9 pageshttpdxdoiorg1011552015376163

2 Journal of Photonics

Figure 1 Machine automated inspection [4]

other one is to use optimal filters such as [18ndash21] In generalfiltering with a filter bank can generate excessive data forprocessing though a set of filters may aid the segmentationCorrespondingly the quality of localization and recognitionis affected dramatically and the time consumption is large aswell

Moreover model based approaches have been success-fully used in the fabric defect detection Some of texturalanalysis methods are based on the Markov random fieldmodel [22] Campbell et al [23] used the model based clus-tering method to detect linear pattern production defects

Recently neural network has attracted a lot of attention indefect detection applicationsHuang andChen [24] have usedback-propagationneural networkwith fuzzy logic to achievethe classification of eight different kinds of fabric defects In[25] an approach for the segmentation of local textile defectsusing feed-forward neural network is presented RecentlyJing et al in [26] present a machine vision system whichuses basic patch statistics from raw image data combinedwith a two-layer neural network to detect surface defects onarbitrary textured and weakly labeled image data

12 Present Work In textured images discontinuities rep-resent generally an important feature such as boundariesIt is known that the local homogeneity analysis is themethod to detect these discontinuitiesMoreover due to theirnonparametric nature and their ability to describe complexdecision regions neural network is one of the best classifiersused for defect detection In this paper a new approachfor fabric defect detection based on the local homogeneityanalysis and neural network is proposed

This paper is organized as follows In Section 2 the localhomogeneity computation (119867-image) for different windowsizes is presented The discrete cosine transform of the 119867-image and feature extraction are then described in Section 3Section 4 provides a system description of the proposedmethod It is followed by the experimental results in Section 5Details of training and test procedures which include ourproposed method for fabric defect detection and detectionperformance are described in Sections 51 and 52 respec-tively Finally Section 6 draws conclusion and summarizesthe paper

Pc

P1

f1

P2

f2

P3f3

P4

f4

P5

f5

P6

f6

P7f7

P8f8

P9f9

P10

P11 P12P13 P14

P15

f15P16

f16

P17

f17

P18

f18

P19

f19

P20

f20

P21

f21

P22

f22

P23

f23

P24

f24

Figure 2 Different pixels used to compute the local homogeneityfor 5 times 5mask for each pixel corresponds to a direction vector 119891

119894

2 Local Homogeneity Definition

It is known that discontinuities in textured images representgenerally an important feature such as boundaries The mainproblem is how to detect these discontinuities A method toresolve this problem is the homogeneity (119867) The calculationof119867 was proposed first by Jing et al in [26] Let (119909 119910) be thecoordinates of a pixel and 119868(119909 119910) the corresponding intensitylevel Let 119875 be a squared window that is a mask of (2119873 + 1)width centered on the pixel In the window 119875 one can definedifferent vectors such as119862119901

119894= (119909119894minus119909119888 119910119894minus119910119888) where (119909

119894 119910119894)

is the neighbor of the center (119909119888 119910119888) defined as

119888 minus 119873 le 119909119894le 119888 + 119873

119888 minus 119873 le 119910119894le 119888 + 119873

(1)

where (119909119894 119910119894) is the coordinate of the pixel 119875

119894 119888 is the

coordinate of the centre pixel (119875119888) and119873 is an integer number

which defines the mask sizeNote that this definition allows computing the homo-

geneity of a pixel for different mask sizes 3 times 3 5 times 5 andso forth

Figure 2 shows an example of the different pixels used tocompute homogeneity in a window of size 5 times 5 pixels

Based on 119862119901119894a new vector 119891

119894is constructed as follows

119891119894= (119868 (119909

119894 119910119894) minus 119868 (119909

119888 119910119888))119862119901119894

10038171003817100381710038171198621199011198941003817100381710038171003817

(2)

For a mask of width (2119873 + 1) we have (2119873 + 1)2 vectors 119891119894

1 le 119894 le (2119873 + 1)2

Let 119891 be the sum of the whole vectors defined in thewindow 119875 so

119891 =

(2119873+1)2

sum

119894=1

119891119894 (3)

The119867 valuemeasurement (the homogeneity value) is definedas the norm of 119891 so that

119867 =10038171003817100381710038171198911003817100381710038171003817 (4)

Journal of Photonics 3

To limit the huge variation of the local homogeneity the nor-malized homogeneity is used

=119867 minus 119867min119867max minus 119867min

(5)

where 119867min and 119867max are the minimum and the maximumvalue of homogeneity respectively The resulting119867-image isconstructed by the different values of computed for eachpixel value 119868

119888(119909119888 119910119888)

3 Discrete Cosine Transform (DCT) andFeatures Extraction from the DCT Block

31 Brief Description of DCT Transform The DCT of an 119867-image119867

119909119910of size119872times119873 is given by (6)This expressionmust

be computed for values of 119906 = 0 1 2 119872 minus 1 and also forV = 0 1 2 119873 minus 1

The variables 119906 and V are frequency variables 119909 and 119910 arespatial variables and 119875 and 119876 are the size of the DCT block

119863119906V

=120590(119906) 120590(V)119875minus1

sum

119909=0

119876minus1

sum

119910=0

119867119909119910

cos((2119910 + 1) 119906120587

2119875) cos((2119909 + 1) V120587

2119876)

(6)

where

120590 (119906) =

radic1

119875 for 119906 = 0

radic2

119875 otherwise

120590 (V) =

radic1

119876 for V = 0

radic2

119876 otherwise

(7)

32 Feature Extraction In order to characterize the fre-quency domain within texture image this paper proposes touse the features energy distribution in the 2D-DCT domainfrom the119867-image which is presented as follows [27]

(i) The horizontal energy value of the DCT block

119864119867=

119875minus1

sum

119906=0

119876minus1

sum

V=0[(V + 1)2 times 1198672

119909119910]12

for 119906 + V = 0 (8)

(ii) The vertical energy value of a DCT block

119864119881=

119875minus1

sum

119906=0

119876minus1

sum

V=0[(119906 + 1)

2times 1198672

119909119910]12

for 119906 + V = 0 (9)

(iii) The diagonal energy value of a DCT block

119864119863=

119875minus1

sum

119906=0

119876minus1

sum

V=0[(119906 + 1) times (V + 1) times 1198672

119909119910]12

for 119906 + V = 0

(10)

0 5 10 15 20 25 300

50

100

150

200

250

300

350

400

Number of features

Eige

nval

ue

Figure 3 Eigenvalues of covariance matrix for feature reduction

(iv) The average energy value of a DCT block

119864119872=

1

119875 times 119876

119875minus1

sum

119906=0

119876minus1

sum

V=01198672

119909119910 (11)

(v) The standard deviation energy value of a DCT block

119864119878=

1

119875 times 119876 minus 1

119875minus1

sum

119906=0

119876minus1

sum

V=0(119867119909119910minus 119864119872)2

(12)

Then the features vector of each pixel is expressed as

119879 = [119864119867 119864119881 119864119863 119864119872 119864119878] (13)

In the next section the principal component analysis(PCA) is introduced to eliminate correlations between fea-tures and reduce the dimensionality of the original featurevectors

33 Features Reduction Feature reductionmeans transform-ing the original features into a lower dimensional space [2829] Most of the feature extraction techniques have beenbased on linear techniques such as principal componentanalysis (PCA) PCA is a quantitatively rigorous methodfor achieving data dimensionality reduction The methodgenerates a new set of variables called principal components(PCs) which maximize the variance of the projected vectorsEach PC is a linear combination of the original variables Allthe PCs are orthogonal to each other so there is no redundantinformation The PCs as a whole form an orthogonal basisfor the space of the data Thus the first PC consists ofthe highest variability the second PC consists of the nexthighest variability and so on for other directions The firstfew components are kept and others with less variabilityare discarded As shown in Figure 3 in the proposed workthe first 4 PCs are used to feed to the feed-forward neuralnetworks (FFN) classifier


Horizontal energy

Diagonal energy

Energy mean

Energy standard deviation

Vertical energy

DCT transform

Feature extraction of DCT blocksH-image FNN classification Input image

EH

ED

EV

EM

ES

Figure 4 Block diagram of fabric defect detection using119867-image and FFN

4 Fabric Defect Detection Method

41 Fabric Defect Classification Based on FFN This methodconsists in scanning the acquired image with a squaredwindow and computing the local homogeneity for each pixeland constructing the 119867-image In our work we have used awindow of size 11 times 11 pixels Then we divided the119867-imageinto overlapping squared blocks The five aforementionedtypes of energy are computed through the DCT applied foreach block which produces these energies Every pixel of theacquired image is characterized by a feature vector (five typesof energy) Then the statistical or 119885-score normalizationtechnique is applied to all the features vectors that form theinput to the neural network The number of the inputs of theneural network is not equal to the same number of featureextractions They are obtained using a principal componentanalyzing (PCA) routine This routine is applied in order tolimit the number of inputs of neural network In this case thenumber of inputs equals four (ie the number of chosen prin-cipal components)The number of the training vectors equalsthe number of pixels in the considered field of view of theinvestigated fabric surface The network has a single output

The concept diagram of the fabric defect detectionscheme is presented in Figure 4

The procedure of the proposed method can be summa-rized as follows

Step 1 Scanning of the input image 119868 with a squared windowand computing the local homogeneity for each pixel

Step 2 Normalization of the local homogeneity

Step 3 Division of the 119867-image into overlapping squaredblocks and application of a DCT transform

Step 4 Features extraction from the DCT block

Step 5 Features reduction using the PCA method

Step 6 Normalization of all features using Zero mean andunity variance technique

Step 7 Classification process for fabric defect by FFN basedon 119867-image (since only 2 conditions (block contains defectregions or not) need to be identified in this paper)

42 Learning Defect Classifier After a feature calculationstep now we use the training data to train a neural networkusing back-propagation An example of a FFN structure isshown in Figure 4 FFN classification The size of the inputlayer is equal to the feature dimension here 119896 = 4 (after apply-ing a PCA technique the number of features is reduced to 4)The second layer called the hidden layer has L1 nodes In theoutput layer all nodes from the hidden layer are mapped toone final node showing the classification result All weightsin the neural network are initialized by random numbersTraining a network means optimizing network detectionperformanceThis is realized byminimizing the performancefunction The performance measurement function in thiscase is the mean squared error (MSE)The final classificationresult is a value between [minus1 1] and predicts whether anunknown block contains defect regions or not

5 Experimental Results Application ofthe Proposed Method in Fabric Defect

This section presents the experimental results for fabric defectdetection The dataset used is established and the results areanalyzed in Section 51 The performance of the proposedalgorithm for a number of sample images is presented andanalyzed in Section 52 Discussion and comparison withsome previous works are given in Section 53

51 Dataset The dataset used in this work is provided byPARTNER textile industry in Tunisia [30] The dataset con-sists of 89 images of which 13 have no defects and 76 havevarious fabric defects The total databases were divided intotwo sets one for training (containing 60 of the samples)and the other for test (containing 40 of the samples) Table 1summarizes some fabric defects types used in this study theirdefinition and their reasons

All simulations were implemented on a personal com-puter (Intel core 2Duo CPU T5870 2GHz and 3GB ofRAM) The programs were tested first in MATLAB languageand after that they were rewritten in C++ language for thepurpose to be implemented in a real time process We notethat once the neural network is trained in the learning phase(offline step) it will be inserted in the defect detection processand then the global needed time is considerably reduced


Table 1 Example of fabric defect types

Defect type Definition Reasons

Floats A portion of a yarn in a fabric that extends or floatsunbound over two or more adjacent ends or picks

It is caused by missing of interlacement of two series ofthreads

Holes A fabric area free of both warp and weft threads It is a mechanical fault caused by a broken machine part

Oil stains Fabric area contains oil spots It is caused by too much oiling on loom parts or fromother external sources

Slubs A local uneven fabric thickness

It is caused by an extra piece of yarn that is woven intofabric It can also be caused by thick places in the yarnor by fly waste being spun in yarn during the spinningprocess

KnotsA fabric place where two ends of yarn have been tiedtogether and the tails of the knot are protruding fromthe surface

It is caused by tying spools of yarn ends together

Miss-end A warp thread is absent in the fabric for a short or longdistance

It is due to incorrect warping or by a broken warpthread that never replaced by another one

Miss-pick A weft thread is absent in the fabric for a short or longdistance

It is caused by incorrect picking or if the weaverrestarted the loom after any stoppage without adaptingthe position for the new insertion

Table 2 Training parameters

Parameter ValueNumber of inputsoutputs 41Number of hidden layers 1Number of hidden layer neurons 9Learning coefficient for hidden layer 06Activation function Bipolar SigmoidInitial weight values (random) 0sim01Learning rule delta DeltaConvergence error 28119890 minus 5

Number of iterations 50120

Training of the network was carried out with the errorback-propagation algorithm by using a set of fabric imagesdata containing 46 images with defect and 8 images withoutdefect covering every possible pattern of defect blocks andnondefect blocks

Initially all the weight values of the neural network wererandomly set to small values between 0 and 01 and thelearning coefficients were set to 06 Several experiments wereconducted for selecting the best FFNThe factors that playeda significant role in the performance of the system were thenumber of hidden layer neurons the learning coefficients ofthe hidden layer neurons and the output layer the activationfunction and the learning rule The parameters of the bestarchitecture for FFN are summarized in Table 2

52 Performance Detection After the FFN is trained with 54images its performance has been tested with the remainingones (35 images) The performance of this algorithm isvalidated by calculating the following performance measurefor the train set and test set separately The classificationaccuracy (CA) is as follows ratio between the total number

Table 3 Definitions of true positive (TP) false positive (FP) truenegative (TN) and false negative (FN) in defect detection

Actually defective Actually defect-freeDetected as defective True positive (TP) False positive (FP)Detected as defect-free False negative (FN) True negative (TN)

of correctly classified test samples and the total number oftest samples Generally detection accuracy also known asdetection success rate is defined as

CA [] = TP + TNTP + FN + TN + FP

times 100 (14)

Table 3 outlines definitions of true positive (TP) falsepositive (FP) true negative (TN) and false negative (FN) indefect detection

The classification results of all classifiers in the trainingand testing processes are presented in Table 4 In the trainingprocess all the FFN classifiers achieve an average accuracyof 9830 and 9905 in the whole dataset and the reduceddataset respectively This indicates that the classifiers arewell trained and can be applied for fabric defect detectionHowever in the testing process these classifiers are validatedagainst the test data the average accuracy is about 9735 and9625 for the reduced and original features respectively

From the test results shown in Table 4 it can be seen thatthe FNN classifiers recognize the blocks defects effectivelyThe recognition results of FNN are ideal because of its highaccuracy and a good generalization capability when theaverage classification efficiency close to 9735 is reasonablygood

For 256 times 256 image the time needed for defect detectionby scanning all the image is about 08 seconds Taking intoaccount the production rate (knitting speed) the algorithmcan be used in a real time application


Table 4 The testing accuracy for different blocks conditions using FNN

Blocks conditionsFNN including all features FNN including selected featuresClassification accuracy [] Classification accuracy []

Training Testing Training TestingDefective blocks 9750 9530 9830 9660Nondefective blocks 9910 9720 100 9810Average 9830 9625 9915 9735

Table 5 Comparison of the proposed method with some previous research

Literature Features Classifier Accuracy []

[29] Complex symmetric Gabor filter bank and principalcomponent analysis (PCA) PCA + Euclidean norm 988

[31] Wavelet based feature extraction and morphologicaloperations + Dempster-Shafer theory MLP neural networks 8948

[32] Radial basis function (RBF) network and gray levelarrangement in the neighborhood of each pixel + PCA PCA + RBF network with Gaussian kernel 834

[33] Small scale overcomplete basis set and Gabor filter Sparse coding 93ProposedMethodology

Features derived from DCT transform of119867-image andFFN PCA + FFN 9735

53 Discussion and Comparison with Some Previous WorksIn this experiment the 119867-image is obtained after scanningthe input image using a window of size 11times11 pixelsThe fea-tures extractions are produced by dividing the119867-image intooverlapping blocks of size 4 times 4 pixels after applying the DCTtransform The final classification result is obtained froma feed-forward network trained with the back-propagation(BP) Figure 5 shows detection examples of the fabric datasetThe kinds of fabric defects vary in size shape and orientation

In Figure 5 column (a) shows the original images Col-umns (b) and (c) respectively correspond to the 119867-imageand defect detection results Image in the first line repre-sents the ldquoholerdquo defect The proposed algorithm was able tocorrectly detect the defect Image in the second line showsanother type of defect (slub defect) and the correspondingdefects detection results which were accurately detectedOther types of sample defects are shown in the rest of thefigure

It is interesting to note that despite the varying size andorientation of the defects the algorithm managed to identifyaccurately most of the defects

Based on these results it can be concluded that the DCTfeatures extraction method effectively improves classificationperformance for the given fabric defect classification prob-lem

Table 5 provides a summary of the studies on automatedfabric defect detection Bissi et al (2013) [29] proposed anapproach for automated texture defect detection in uniformand structured fabrics based on a complex symmetric Gaborfilter bank and PCAThe performance of their algorithm hasbeen extensively evaluated by using images of the TILDATextile Texture Database and reported an accuracy of 988Tabassian et al (2011) [31] used a wavelet based featureextraction method beside binary thresholding technique and

morphological operations and classified them byMLP neuralnetworks The Dempster-Shafer theory of evidence was thenemployed for combining different evidence to decrease thevalue of uncertainty It achieved a detection accuracy of8948 Zhang et al (2010) [32] used the neighborhood ofeach pixel to extract the features and to reduce them usingPCA and classified them using radial basis function (RBF)network improved by Gaussian mixture model (GMM) withan accuracy of 834 Qiuping et al (2014) [33] used anapproach for defect detection applied to twill plain ginghamand striped fabric using Gabor filter and sparse coding withan accuracy of 93

In conclusion the evaluation of our method versus thosein previous researches given in Table 5 shows that ourexperiment has good potential for improving fabric defectdetection The features derived from the DCT transform of119867-imagewith the FFN classifier produce a good classificationperformance the average CA result is 9735 This combi-nation of the feature and classifier is promising for a highaccuracy of fabric defect detection

6 Conclusions and Future Work

In this paper a novel method is proposed to solve theproblem of fabric defect detection which is based on the localhomogeneity analysis and the feed-forward neural networksA DCT transform based feature extraction method besidethe PCA was used The performance of this algorithm hasbeen extensively evaluated by using images of the PARTNERTextile Texture DatabaseThe comparison with other existingalgorithms reported in literature highlights the effectivenessof the proposed approach The average accuracy of blockcondition was 9735 the results show that the proposedalgorithm is a strong technique for fabric defect detection


Hol

eSl

ubKn

otFl

oat

Miss

-pic

kM

iss-e

nd

(a) (b) (c)

Figure 5 Defect detection results (a) Images of fabrics with different types of defects (b) the corresponding 119867-image and (c) the cor-responding defect detection results


Thealgorithmpresents somedrawbacks namely the arbitrar-ily choice of the windows size of the mask used to computethe 119867-image This in its turn affects intensively the defectdetection of the algorithm Indeed if the mask size is chosenrelatively large (such as 15 times 15) this will homogenize evensome different regions and then this will cause some falsealarms If the mask size is chosen relatively small (such as3 times 3 or 5 times 5) this will induce a poor homogenization andtherefore similar areas (especially in stochastic textures) willbe considered as different areas andmay be seen as defects bythe algorithm

Conflict of Interests

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

Acknowledgment

The authors would like to thank Mr Mohamed Ali DerbelDirector of the PARTENERTextile for helpful discussion andfor providing the fabric image database

References

[1] P Sengottuvelan A Wahi and A Shanmugam ldquoAutomaticfault analysis of textile fabric using imaging systemsrdquo ResearchJournal of Applied Sciences vol 3 pp 26ndash31 2008

[2] T Vikrant and S Gaurav ldquoAutomatic fabric fault detection usingmorphological operations on bit planerdquo International Journal ofEngineering Research amp Technology vol 2 pp 856ndash861 2013

[3] Hong Kong Productivity Council Textile Handbook 2000 TheHong Kong Cotton Spinners Association 2000

[4] H Y T Ngan G K H Pang and N H C Yung ldquoAutomatedfabric defect detectionmdasha reviewrdquo Image andVision Computingvol 29 no 7 pp 442ndash458 2011

[5] M Ghazvini S A Monadjemi N Movahhedinia and KJamshidi ldquoDefect detection of tiles using 2D-wavelet transformand statistical featuresrdquoWorld Academy of Science Engineeringand Technology vol 37 pp 901ndash904 2009

[6] R W Conners C W McMillin and K Lin ldquoIdentifying andlocating surface defects in wood part of an automated lumberprocessing systemrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 5 no 6 pp 573ndash583 1983

[7] D Zhu R Conners and P Araman ldquoCT image sequenceprocessing for wood defect recognitionrdquo in Proceedings of theIEEE 23rd Southeastern Symposium on System Theory pp 75ndash79 Columbia SC USA 1991

[8] C Boukouvalas ldquoColor grading of randomly textured ceramictiles using color histogramsrdquo IEEE Transactions on IndustrialElectronics vol 46 no 1 pp 219ndash226 1999

[9] K Wiltschi A Pinz and T Lindeberg ldquoAutomatic assessmentscheme for steel quality inspectionrdquoMachine Vision and Appli-cations vol 12 no 3 pp 113ndash128 2000

[10] C-H Chan and G K H Pang ldquoFabric defect detection by Fou-rier analysisrdquo IEEE Transactions on Industry Applications vol36 no 5 pp 1267ndash1276 2000

[11] X Yang G Pang and N Yung ldquoRobust fabric defect detec-tion and classification using multiple adaptive waveletsrdquo IEE

ProceedingsmdashVision Image and Signal Processing vol 152 pp715ndash723 2005

[12] D M Tsai and B Hsiao ldquoAutomatic surface inspection usingwavelet reconstructionrdquo Pattern Recognition vol 34 no 6 pp1285ndash1305 2001

[13] Y X Zhi G K H Pang and N H C Yung ldquoFabric defectdetection using adaptive waveletrdquo in Proceedings of the IEEEInterntional Conference on Acoustics Speech and Signal Process-ing (ICASSP rsquo01) pp 3697ndash3700 May 2001

[14] YHan andP Shi ldquoAn adaptive level-selectingwavelet transformfor texture defect detectionrdquo Image and Vision Computing vol25 no 8 pp 1239ndash1248 2007

[15] A Latif-Amet A Ertuzun and A Ercil ldquoAn efficient methodfor texture defect detection sub-band domain co-occurrencematricesrdquo Image and Vision Computing vol 18 no 6 pp 543ndash553 2000

[16] A Kumar and G K H Pang ldquoFabric defect segmentation usingmultichannel blob detectorsrdquoOptical Engineering vol 39 no 12pp 3176ndash3190 2000

[17] A Kumar and G K H Pang ldquoDefect detection in texturedmaterials using Gabor filtersrdquo IEEE Transactions on IndustryApplications vol 38 no 2 pp 425ndash440 2002

[18] A Hamid A Alireza and S Esmaeil ldquoDefect detection intextiles usingmorphological analysis of optimalGaborwaveletrdquoin Proceedings of the International Conference on Computer andAutomation Engineering (ICCAE rsquo09) pp 26ndash30 2009

[19] A Bodnarova M Bennamoun and S Latham ldquoOptimal gaborfilters for textile flaw detectionrdquo Pattern Recognition vol 35 no12 pp 2973ndash2991 2002

[20] R Han and LM Zhang ldquoFabric defect detectionmethod basedonGabor filtermaskrdquo in Proceedings of theWRIGlobal Congresson Intelligent Systems (GCIS rsquo09) pp 184ndash188 Xiamen ChinaMay 2009

[21] M Li and R C Staunton ldquoOptimum Gabor filter design andlocal binary patterns for texture segmentationrdquo Pattern Recog-nition Letters vol 29 no 5 pp 664ndash672 2008

[22] S Ozdemir and A Ercil ldquoMarkov random fields and Kar-humen-Loeve transforms for defect inspection of textile prod-uctsrdquo in Proceedings of the IEEE Conference on Emerging Tech-nologies and Factory Automation vol 2 pp 697ndash703 1996

[23] J G Campbell C Fraley F Murtagh and A E Raftery ldquoLinearflaw detection in woven textiles using model-based clusteringrdquoPattern Recognition Letters vol 18 no 14 pp 1539ndash1548 1997

[24] C-C Huang and I-C Chen ldquoNeural-fuzzy classification forfabric defectsrdquo Textile Research Journal vol 71 no 3 pp 220ndash224 2001

[25] A Kumar ldquoNeural network based detection of local textiledefectsrdquo Pattern Recognition vol 36 no 7 pp 1645ndash1659 2003

[26] F Jing M Li H-J Zhang and B Zhang ldquoUnsupervised imagesegmentation using local homogeneity analysisrdquo in Proceedingsof the IEEE International Symposium on Circuits and Systems(ISCAS rsquo03) pp II456ndashII459 IEEE May 2003

[27] Y-S Chiu and H-D Lin ldquoAn innovative blemish detection sys-tem for curved LED lensesrdquo Expert Systems with Applicationsvol 40 no 2 pp 471ndash479 2013

[28] S Ding C Li and Z Liu ldquoFabric defect detection scheme basedon Gabor filter and PCArdquo Advanced Materials Research vol482ndash484 pp 159ndash163 2012

[29] L Bissi G Baruffa P Placidi E Ricci A Scorzoni and P ValigildquoAutomated defect detection in uniform and structured fabricsusing Gabor filters and PCArdquo Journal of Visual Communicationand Image Representation vol 24 no 7 pp 838ndash845 2013


[30] 2013 httpwwwpartnertextilecom[31] M Tabassian R Ghaderi and R Ebrahimpour ldquoKnitted fabric

defect classification for uncertain labels based on Dempster-Shafer theory of evidencerdquo Expert Systems with Applicationsvol 38 no 5 pp 5259ndash5267 2011

[32] Y Zhang Z Lu and J Li ldquoFabric defect classification usingradial basis function networkrdquo Pattern Recognition Letters vol31 no 13 pp 2033ndash2042 2010

[33] Z Qiuping W Minyuan L Jie and D Dexiang ldquoFabricdefect detection via small scale over-complete basis setrdquo TextileResearch Journal vol 84 no 15 pp 1634ndash1649 2014

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

High Energy PhysicsAdvances in

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


FluidsJournal of

Atomic and Molecular Physics

Journal of



Advances in Condensed Matter Physics

OpticsInternational Journal of



AstronomyAdvances in

International Journal of


Superconductivity


Statistical MechanicsInternational Journal of


GravityJournal of


AstrophysicsJournal of


Physics Research International


Solid State PhysicsJournal of

Computational Methods in Physics

Journal of



Soft MatterJournal of

Hindawi Publishing Corporationhttpwwwhindawicom

AerodynamicsJournal of

Volume 2014


PhotonicsJournal of


Journal of

Biophysics


ThermodynamicsJournal of


Figure 1 Machine automated inspection [4]

other one is to use optimal filters such as [18ndash21] In generalfiltering with a filter bank can generate excessive data forprocessing though a set of filters may aid the segmentationCorrespondingly the quality of localization and recognitionis affected dramatically and the time consumption is large aswell

Moreover model based approaches have been success-fully used in the fabric defect detection Some of texturalanalysis methods are based on the Markov random fieldmodel [22] Campbell et al [23] used the model based clus-tering method to detect linear pattern production defects

Recently neural network has attracted a lot of attention indefect detection applicationsHuang andChen [24] have usedback-propagationneural networkwith fuzzy logic to achievethe classification of eight different kinds of fabric defects In[25] an approach for the segmentation of local textile defectsusing feed-forward neural network is presented RecentlyJing et al in [26] present a machine vision system whichuses basic patch statistics from raw image data combinedwith a two-layer neural network to detect surface defects onarbitrary textured and weakly labeled image data

12 Present Work In textured images discontinuities rep-resent generally an important feature such as boundariesIt is known that the local homogeneity analysis is themethod to detect these discontinuitiesMoreover due to theirnonparametric nature and their ability to describe complexdecision regions neural network is one of the best classifiersused for defect detection In this paper a new approachfor fabric defect detection based on the local homogeneityanalysis and neural network is proposed

This paper is organized as follows In Section 2 the localhomogeneity computation (119867-image) for different windowsizes is presented The discrete cosine transform of the 119867-image and feature extraction are then described in Section 3Section 4 provides a system description of the proposedmethod It is followed by the experimental results in Section 5Details of training and test procedures which include ourproposed method for fabric defect detection and detectionperformance are described in Sections 51 and 52 respec-tively Finally Section 6 draws conclusion and summarizesthe paper

Pc

P1

f1

P2

f2

P3f3

P4

f4

P5

f5

P6

f6

P7f7

P8f8

P9f9

P10

P11 P12P13 P14

P15

f15P16

f16

P17

f17

P18

f18

P19

f19

P20

f20

P21

f21

P22

f22

P23

f23

P24

f24

Figure 2 Different pixels used to compute the local homogeneityfor 5 times 5mask for each pixel corresponds to a direction vector 119891

119894

2 Local Homogeneity Definition

It is known that discontinuities in textured images representgenerally an important feature such as boundaries The mainproblem is how to detect these discontinuities A method toresolve this problem is the homogeneity (119867) The calculationof119867 was proposed first by Jing et al in [26] Let (119909 119910) be thecoordinates of a pixel and 119868(119909 119910) the corresponding intensitylevel Let 119875 be a squared window that is a mask of (2119873 + 1)width centered on the pixel In the window 119875 one can definedifferent vectors such as119862119901

119894= (119909119894minus119909119888 119910119894minus119910119888) where (119909

119894 119910119894)

is the neighbor of the center (119909119888 119910119888) defined as

119888 minus 119873 le 119909119894le 119888 + 119873

119888 minus 119873 le 119910119894le 119888 + 119873

(1)

where (119909119894 119910119894) is the coordinate of the pixel 119875

119894 119888 is the

coordinate of the centre pixel (119875119888) and119873 is an integer number

which defines the mask sizeNote that this definition allows computing the homo-

geneity of a pixel for different mask sizes 3 times 3 5 times 5 andso forth

Figure 2 shows an example of the different pixels used tocompute homogeneity in a window of size 5 times 5 pixels

Based on 119862119901119894a new vector 119891

119894is constructed as follows

119891119894= (119868 (119909

119894 119910119894) minus 119868 (119909

119888 119910119888))119862119901119894

10038171003817100381710038171198621199011198941003817100381710038171003817

(2)

For a mask of width (2119873 + 1) we have (2119873 + 1)2 vectors 119891119894

1 le 119894 le (2119873 + 1)2

Let 119891 be the sum of the whole vectors defined in thewindow 119875 so

119891 =

(2119873+1)2

sum

119894=1

119891119894 (3)

The119867 valuemeasurement (the homogeneity value) is definedas the norm of 119891 so that

119867 =10038171003817100381710038171198911003817100381710038171003817 (4)




(5)


119888(119909119888 119910119888)






119863119906V

=120590(119906) 120590(V)119875minus1

sum

119909=0

119876minus1

sum

119910=0

119867119909119910

cos((2119910 + 1) 119906120587

2119875) cos((2119909 + 1) V120587

2119876)

(6)

where

120590 (119906) =

radic1

119875 for 119906 = 0

radic2

119875 otherwise

120590 (V) =

radic1

119876 for V = 0

radic2

119876 otherwise

(7)



119864119867=

119875minus1

sum

119906=0

119876minus1

sum

V=0[(V + 1)2 times 1198672

119909119910]12

for 119906 + V = 0 (8)


119864119881=

119875minus1

sum

119906=0

119876minus1

sum

V=0[(119906 + 1)

2times 1198672

119909119910]12

for 119906 + V = 0 (9)


119864119863=

119875minus1

sum

119906=0

119876minus1

sum

V=0[(119906 + 1) times (V + 1) times 1198672

119909119910]12

for 119906 + V = 0

(10)

0 5 10 15 20 25 300

50

100

150

200

250

300

350

400

Number of features

Eige

nval

ue



119864119872=

1

119875 times 119876

119875minus1

sum

119906=0

119876minus1

sum

V=01198672

119909119910 (11)


119864119878=

1

119875 times 119876 minus 1

119875minus1

sum

119906=0

119876minus1

sum

V=0(119867119909119910minus 119864119872)2

(12)


119879 = [119864119867 119864119881 119864119863 119864119872 119864119878] (13)




Horizontal energy

Diagonal energy

Energy mean


Vertical energy

DCT transform


EH

ED

EV

EM

ES











































times 100 (14)


























Hol

eSl

ubKn

otFl

oat

Miss

-pic

kM

iss-e

nd

(a) (b) (c)






Acknowledgment


References









































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics






(5)


119888(119909119888 119910119888)






119863119906V

=120590(119906) 120590(V)119875minus1

sum

119909=0

119876minus1

sum

119910=0

119867119909119910

cos((2119910 + 1) 119906120587

2119875) cos((2119909 + 1) V120587

2119876)

(6)

where

120590 (119906) =

radic1

119875 for 119906 = 0

radic2

119875 otherwise

120590 (V) =

radic1

119876 for V = 0

radic2

119876 otherwise

(7)



119864119867=

119875minus1

sum

119906=0

119876minus1

sum

V=0[(V + 1)2 times 1198672

119909119910]12

for 119906 + V = 0 (8)


119864119881=

119875minus1

sum

119906=0

119876minus1

sum

V=0[(119906 + 1)

2times 1198672

119909119910]12

for 119906 + V = 0 (9)


119864119863=

119875minus1

sum

119906=0

119876minus1

sum

V=0[(119906 + 1) times (V + 1) times 1198672

119909119910]12

for 119906 + V = 0

(10)

0 5 10 15 20 25 300

50

100

150

200

250

300

350

400

Number of features

Eige

nval

ue



119864119872=

1

119875 times 119876

119875minus1

sum

119906=0

119876minus1

sum

V=01198672

119909119910 (11)


119864119878=

1

119875 times 119876 minus 1

119875minus1

sum

119906=0

119876minus1

sum

V=0(119867119909119910minus 119864119872)2

(12)


119879 = [119864119867 119864119881 119864119863 119864119872 119864119878] (13)




Horizontal energy

Diagonal energy

Energy mean


Vertical energy

DCT transform


EH

ED

EV

EM

ES











































times 100 (14)


























Hol

eSl

ubKn

otFl

oat

Miss

-pic

kM

iss-e

nd

(a) (b) (c)






Acknowledgment


References









































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics




Horizontal energy

Diagonal energy

Energy mean


Vertical energy

DCT transform


EH

ED

EV

EM

ES











































times 100 (14)


























Hol

eSl

ubKn

otFl

oat

Miss

-pic

kM

iss-e

nd

(a) (b) (c)






Acknowledgment


References









































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics




























times 100 (14)


























Hol

eSl

ubKn

otFl

oat

Miss

-pic

kM

iss-e

nd

(a) (b) (c)






Acknowledgment


References









































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics
























Hol

eSl

ubKn

otFl

oat

Miss

-pic

kM

iss-e

nd

(a) (b) (c)






Acknowledgment


References









































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics







Acknowledgment


References









































FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics













FluidsJournal of


Journal of










Superconductivity




GravityJournal of








Journal of






Volume 2014


PhotonicsJournal of


Journal of

Biophysics



research article fabric defect detection using local...

Documents