Image Analysis of Fabric Pilling Based on Light Projection

Journal of Donghua University (Eng. Ed.) Vol. 20, No.4 (2003)Image Analysis of Fabric Pilling Based on Light ProjectionCHEN Xia (陈霞) ,HUANG Xiu-bao(黄秀宝)College of Textiles, Donghua University, Shanghai, 200051, ChinaThe objective assessment of fabric pilling based on lightimages of pilled fabric samples. The fabric samples areprojection and image analysis has been exploited recently.fixed in advance on a conveyer belt driven by a motor andThe device for capturing the cross- sectional images of thecan be wrapped closely around a shaped bar which ispilled fabrics with light projection is elaborated. Thespecially designed to be characteristic of large curvature indetection of the profile line and integration of the sequentialthe bar edge. When the sample arrived the edge of the barcross- sectional pilled image are discussed. The thresholdwhere a light source was positioned below， sequentialbased on Gaussian model is recommended for pillprojected cross-sectional images of the pilled fabric samplessegmentation. The results show that the installed system iswere captured by a CCD camera which was positionedcapable of eliminating the interference with pill informationabove the bar. And those projected images were then readfrom the fabric color and pattern.into computer to be analyzed.Keywords: fabric pilling， image analysis of fabricpilling，light projection， image, edge detection ofprojected profile, pill segmentationhe plling of textile fabric is a well- known phenomenon.\29The development of pills on a fabric surface, in addition tospoiling the original appearance and poor feeling, initiates the8attrition of garment and reduces the serviceability. Pllingresistance is usually tested in textile trade and quality control1: Fabric sample 4: Shaped bur 7: CCD camerawhere the fabric sample is tumbled or treated with abrasive2: Coaveyer belt 5: Light source 8: Tension regulating rllematerials that simulate the normal wear. Then the pilled3: Motor driver 6: Ground glass 9: Cabinetig. 1 Image capturing configurationsample is compared with the photographs of standard pilledIn this research, the size of projected image is 400X 50fabric samples by visual inspection to evaluate the pilling degreepixels with the resolution 0. 15 mm/ pixel. And the intervalthat usually ranges from grade 5 (no plling) to grade 1 (veryof step motion is also 0. 15 mm. The total number ofsevere pilling). As the inconsistency and the inaccuracy of theprojected images for each sample is 400 frames， so therating results obtained using visual method, more reliable andactual field analyzed by this system is 60X 60 mm’in size .objective methods for pilling evaluation are desirable for the .on each sample surface.textile industry.Many textile scientists have made valuable contributions tothe objective assessment of fabric piling [-1， All these .Edge detection of Profile Lineresearch work， however, has the common difficulty inFigure 2 (a) shows a projected image of the pilledeliminating the interference with the fabric pill informationfabric captured by light projection method, its obect linefrom the fabric color and pattern, which will finally affectthe pill ratings. A new method based on light projection iskeeping the fabric separate from the background reflectsthe pilling information at this section. As image is alwaysproposed, which may be the way to overcome the problemnoised randomly, the edge line of each projected image wasmentioned above. As a part of the work, the design anddetected with LOG ( Laplacian of Guassian) operatorinstallation of the device for capturing the images of pilledfabric samples， the processing of these images and thewhich can smooth the noises. The function of LOG filter isa second differential operator given bysegmentation of pills on images are presented in this paper.a'G，a"G__ 1中国煤化工丽( 2o-)Device for Capturing Projected ImagesTYHCNMHG+2\(1)Figure 1 shows the device for capturing the projected*Received Mar. 21, 2003This research was supported by the Research Fund for Doctoral Program of Higher Education (No. 99025508)Correspondence should be addressed to HUANG Xiu-bao, professorJournal of Donghua University (Eng. Ed.) Vol. 20，No. 4 (2003)(4)6Sample No.1Sample No.2Sample No 3Fig. 2 Projected image (a)， convolution image (b) and result ofedge detection (c) of a fabric sampleThe projected image was convolved with LOG filter,then the edge line can be found by detecting all such pointson the convolved image as shown in figure 2(b) that theconvolved values of the image change their sign frompositive to negative or from negative to positive at theseSample No .4Sample No .5points. The scale σ of the LOG operator should beFig.4 Images of five fabric samplesreasonably selected according to the noise intensity andvariation range of the gray level at the image edge. AsPreprocessing of Integrated Imagesscale σ increases, the operator is capable of better anti-interference with the noise， but the edge details in theThe images obtained should be further preprocessedregion less than 2√2σ will be probably missed.for many reasons. As mentioned above the edge detailsThe white curve shown in figure 2(c) is the detectedwith size less than 2V2σ will be probably missed when σedge profile line of the image shown in figure 2(a). Inincreases, it will sometimes cause a gap on the sectional pillorder to make the pills information operated digitally, theprofile. The image background is usually uneven becauseheight value of each white pixel is further extracted withthe sample conveying system can not be in motion withoutlinear interpolation and using pixel as height unit. Figure 3even very little instability and it is also difficult to mountshows the extracted height variation of the edge profilethe shaped bar absolutely leveled. In some cases， theline, which is actually the digital pills information of onesurface of the sampled fabric is randomly uneven, whichsectional pilled fabric. Integrating all the edge profiles ofwill also result in the image background uneven. It is,400 frames projected images, a three-dimension profiletherefore，necessary to remove all these undesirable factors .surface of pilled fabric was obtained. The correspondingaffcting the image background by using proper imagetwo-dimensional gray image of pilled samples can bepreprocessing.developed by converting the height value matrix of th(1) The elimination of pill gapssurface into gray scale one. Five gray images of pilledThe gaps on the sectional image of pill profile arefabric samples are obtained as shown in figure 4 where thesimilar to the image noise of negative pulse type and can beinterference with the fabric pill information from theremoved by the closing operation of mathematical .fabric color and pattern is eliminated, because these graymorphologyimages originate from projected cross-sectional images. InThe closing operation is the process of gray dilationthe figures, the height value in the bright region is higherfollowed by gray erosion and defined asthan that in adjacent dark region, which means there is a. f.b= (fθb6)⑧b(2)pill present in the bright region.25 fwhere b is a structure element, fθb is the operation ofgray dilation, fb, gray erosion.。2(Figure 5 shows an original pill gap，plotted using15dark line and dot, but it is properly renovated with the中国煤化itted line and symbolThes three-by three matrix100200300400as::TYHCNMHGposition(column)Fig.3 Height of profile linecTherefore the closing operation has the effect offilling small and thin holes in objects， connecting theJournal of Donghua University (Eng. Ed.) Vol. 20, No.4 (2003)adjacent objects， such as the divided pills on the pilledbackground unevenness has been obviously improved.fabric image，and generally smoothing the boundaries oflarger objects without changing their area.16 [12Sample No.ISample No.2Sample No.3original dataresult of closingSample No.4Sample No.51S60165Fig. 6 Preprocessed imagesposition(column)Fig.5 the effect of eliminating pil holePill Segmentation(2) The elimination of image bac kground unevennessThe unevenness of the image background， whichAfter the pilled fabric image being preprocessed, itcomes from the unlevelling mount of the shaped bar,should be further segmented to separate the pill targetusually causes the image gray changing tendentiously asfrom the fabric background. Figure 7(a) and 7(b) showshown in Fig 4 where the image on the left side is brighterthe gray histogram and Gaussian model fitted line ofthan that on the right side and each image appears to be thepilled fabric images with pilling grade 5 and pilling gradesame unevenness. It can be improved by subtracting al respectively. The pilled fabric with grade 5 isspecial matrix from the original image matrix. The specialcharacterized by no pill on its surface and its image graymatrix is such a one that the element value is constanthistogram reflects essentially the gray variation of thwithin each individual column, but for different columns,imagebackground. Meanwhile, the fitted line as shown init varies linearly according to the changing tendency of the .figure 7(a) for the image gray of pilled fabric with gradegray level of the original image.5 coincides very well with the original gray histogram,The unevenness caused by the instable movement ofwhich means that the background gray of the pilled imagethe sample will result in the gray level variation of thecan be simulated using Gaussian model. In figure 7 ( b)image background one by one between the columns. It canthere is a lttle difference between the fitted line andbe eliminated with the same method as above, but thehistogram on the right tail region，this is because of thematrix subtracted from the original image matrix should bepresence of many pills on the pilled fabric with gradethe other one， in which the element value is constantHowever, in the peak region the Gaussian model fitting iswithin each individual row，but it changes for different .still perfect. For the fact discussed above a thresholdrows depending on the original image gray level changingbased on Gaussian model is recommended to segment thealong the columns.pills from the image background.The random unevenness of the sampled fabric surfacewill make the background image brighter in some regions8 000>ut darker in other regions without any order. This6000unevenness can be removed by Top-Hat transformation ofHistogrammathematical morphology as indicated in equation (3)[12] 。4000Fitted line .h= f-(f●b)3)where f is the input image, h is the output image, f●b is中国煤化工opening operation, which is defined as the gray erosionMYHCNMHG- 101sofollowed by gray dilation. The size of the structure elementgray valueshould be bigger than that of the biggest pill.(a)Figure 6 shows the results of the pilled fabric imageafter preprocessing. In the figure, it can be seen that theJournal of Donghua University (Eng. Ed.) Vol. 20，No. 4 (2003)grade 1 to grade 5. On the sample labeled grade 1 there is8000severe pilling, but for grade 5 no pilling. From the figures6000it can be seen that the effect of pill segmentation with theHistogramthreshold based on Gaussian model is obviously satisfied4000....Fitted linewith the fact that the pill number and pill area decreasegradually from sample grade 1 to grade 5.2000Conclusion05010015gray valueThe device for capturing the projected pilled images is an(b)innovational one，because it is capable of eliminating theFig.7 Gaussian ftting for histograminterference with the pill information of the captured imagesThe Gaussian model is expressed as the follows:from the fabric color and pattern. The experimented results(x- u)2validate that all the methods used are effective for detectingg=Aexp(- 2dthe profile line of projected images and preprocessing'The threshold U is given by:transferred two -dimensional images. Also the threshold basedon Gaussian model is worthwhile to recommend for segmentingU= u+Nσpills from fabric background.where N is a real positive parameter. For different valueof N, the probability of the background pixels being mis-Referencesclassified as pills is varied. When the value of N is taken tobe2，2.5，3，3.5，4, 5 the corresponding values of the[1] Konda A., Xin L.，etc.，Journal of the Textile Machinerymis-classified probability are 2. 28%，0. 27%，0. 025%，Society of Japan, 1990，36(3)， 96- 1070.003% and 0. 000 003%。[2] Ramgulam R. B.. Amirbayat J.. and Porat I.. Journal ofFigure 8 shows the results of pill segmentation for fivethe Textile Institute, 1993, 84(3)， 221 - 226samples with pilling grade ranging 1 to 5. On these images,[3] Ramgulam R. B.. Amirbayat J.. and Porat I.. Journal ofthe white region is the pill area and the dark part is thethe Textile Institute, 1993, 85(4), 397 - 401fabric background. According to the national standards ，[4] His C. H.，Bresee R. R.，etc.， Journal of the TextileInstitute, 1998, 89(1), 80-95the pilled fabric samples are classified as five grades from[5] His C. H，Bresee R. R.，etc.. Journal of the TextileInstitute, 1998. 89(1)， 80-95[6] Xu. B.，Journal of the Textile Institute, 1997， 88(4)，[7] Abril H.C., Millan M.S., and Navarro R.，1996， Proc.SPIE, 2786，19 - 28( 1996)[8] Abril H. C.. Millan M.S., etc., 1997， Proe. SPIE, 3101.Sample gradelSample grade2Sample grade 3283 - 291(1997)[9] Abrid H.C.，Torres Y. And Navarro R.. 1998, Opl Eng，3711), 2937 -2947[10] Abril H.C.. Millan M.S. And Torres Y.，2000. Opt Eng ,39(6)，1477- 1487[11] Airikasemlert A. And Tao x.，Textile Research of Journal, .2000，70(12), 1076- 1087[12] 龚炜,石青云,程民德.数字空间中的数学形态学. 北京,科学出版社,1996, 8-10Sample grade 4Sample grade5[13] 王润生,图像理解，长沙,国防科技大学出版社，1994，Fig. 8 Segmented images for five samples93-101中国煤化工MHCNMHG