织物疵点检测

DEFECT DETECTION IN REPETITIVE FABRIC PATTERNSR.PerezComputer Engineering Technical University of Valencia Ferrandiz y Carbonell no.2Alcoy,Spain. email:ruperez@disca.upv.esJ.SilvestreComputer EngineeringTechnical University of Valenciaemail:jsilves@disca.upv.esJ.MunozInstituto Tecnologico TextilEmilio Sala no.1Alcoy,Spainemail:jmunoz@aitex.esABSTRACTIn this paper a automated analysis system for defect de-tection in the print process offlocked fabrics with repet-itive patterns is presented.This process represents a less computationally complex method than the detection of any type of pattern in the print process.For this reason,this problem can be solved using a personal computer(PC)in-specting100%of the production that fulfill the require-ments of repetitive patterns.This system will be extremely useful for companies which produce repetitive pattern tex-tiles in large quantities.The implementation of a prototype system,based on the Fast Fourier Transform,will be de-scribed.KEY WORDSTextile Inspection,Fast Fourier Transform1IntroductionIn spite of the advantages that computer vision provides to the fabric inspection process,in the production stage as well asfinal inspection,this technology has had little impact on the textile sector when considering developed systems,research articles and research institute projects. Meier[1]points out that although no electronic system is equal to the human eye,there are disadvantages.These include concentration,reproducibility of the process,es-pecially with the entire width of the fabric,speed and spe-cial functions:analyzing the size,form,configuration,con-trast and relevancy of the defect to thefinality of the prod-uct.However,the textile industry employs computer vision very little because:•Visual inspection of irregularities is complex due to the fact that the product is sometimes irregular by na-ture.The textile process can vary significantly and it can still be considered correct.•Fabric materials areflexible and can easily be dis-torted,which makes it difficult to use standard meth-ods for automated inspection.•The economic cost it is possibly greater in this sec-tor than in others,because of the size of the area to be inspected and the resolution required to detect the defects.For these reasons,the development of systems governed by these restrictions,capable of detecting errors visible to the human eye,are quite expensive.Nevertheless,there have been serious attempts by researchers to introduce these techniques in the textile sector.For example,in regular texture fabrics analysis,Shin[2]develops an algorithm for defect detection and classification based on energy mea-sures and the adaptive mask theorem.Although this needs the design of a classifier using test data,it is not a gener-ally accepted method(Campbell[3-6])due to the impos-sibility of obtaining representative defect data with all er-ror characteristics.Cohen[7]develops a system based on a Gaussian Markov Random Field(GMRF)model of the correct textile texture,using non-overlapped windows to extract the GMRF parameters and compare it using a clas-sifier,with the correct characteristics of the fabric texture to determine the correctness of the window.Other work on texture defect detection can be seen in papers by Camp-bell[3-6]and Balakrishnan[8],based on Fourier Trans-form and where they also offer an interesting description of the fabric problems,Hoffer[9]uses an Optic Fourier Trans-form(OFT),Lewis[10],develops a Fourier based method using multiresolution decomposition by wavelets,and Es-cofet[11]uses the Fourier Angular Spectra Correlation to detect global defects and Gabor Filters to detect local de-fects.This paper is organized as follow:In section2we explain the problems and their particularities,in section3 we present the developed algorithm and in section4the results and the conclusions are shown.2Problem descriptionIn the textile industry,the print process is the way in which a particular substance,usually dye,is impregnated on to a fabric in some areas using a perforated cylinder with holes positioned to give the desired design.A cylinder is used for each colour in the process.The most common problem in this process is the obstruction of the cylinder holes by dust, solidification of the dye,etc.so that the resulting design is faulty.In theflocked fabrics industry,it is normal proce-dure to use this process to makeflocked drawings on fabric bases.The colour of the fabric bases is usually the same as theflock to be printed,as well as the adhesive used tofix theflock to the weave.Therefore,although this process ofproduction is usually slower that the normal fabric printing processes and only works with one colour,it presents a se-rious disadvantage to the inspection of defects by the work-ers.This disadvantage is that all the elements used have the same colour,which makes detection very difficult.When considering that up to10,000meters of the same design and colour can be produced,it is easy to see the importance of automatic detection of the printing defects.The system has to detect the defects produced in the print process of the adhesive on the fabric bases,so that later theflock adheres to the desired areas.The system designed does not require hi-res images since the geometric patterns are not exactly equal.In addition,when the adhesive fails in a small area, the adhesion process masks this so it is imperceptible in thefinished product.Another factor introduced by the im-perfection of the textile geometric designs and what it is tried to avoid in the stamping process when the defect is detected,it will be signalled to the operator,who can clean the cylinder in the area indicated.Thus,the defect is cor-rected and the number of defective meters is minimized. 3Approach3.1HardwareFor the detection of defects,a colour linear camera of2048 pixels has been used,which provides sufficient resolution for a fabric analysis of1,60meters.For the illumination system,due to the opacity of the fabric base,a front light system need to be used,that provides an adequate contrast between the fabric and adhesive,obtaining therefore im-ages where the design can be appreciated.3.2Developed algorithmThe algorithm used is based on the calculation of the fast Fourier Transform(FFT)on windows throughout all the image.To each one of these windows the forward FFT is applied,and then frequencies arefiltered using a mask on the transformed image.Once the image isfiltered,the inverse transform is applied to obtain thefiltered real im-age in which the defects appear as a white or dark spots, as thefiltering process removes the repetitive pattern.In a post process stage,the repetition of these defects is ana-lyzed since those that are repeated in all rapports constitute significant defects for the company.3.2.1Windowing processThe FFT window size is set to256pixels.Different win-dow sizes were tested and the best results were obtained using this window size with the resolution used.To con-struct the windows,a superimposition among them has been made so that only the central part of the inversed im-ages is processed,avoiding edge problems on the FFT win-dow.For the overlapping of the different FFT windows,it has been taken into account that the quantity of windows to process covers the whole image plus an external border similar to the overlapping which exists between the rest of the windows.This external border added to the original image avoids border problems in the margins of the image. This extra border which together with the rest of the image we will call amplified image,isfilled with the inverted im-age which we analyse.Infigure1the windowing process can be seen with the overlapping of thewindows.Figure1.FTT Windows on aflock adhesive image.The contrast has been exaggerated3.2.2Windows processingOnce the windowing process of the image is calculated,the FFT is done on each ter,the necessary calcu-lations are made to obtain thefilter mask that will elimi-nate the frequencies whose module is superior to a certain threshold value,which will belong to the repetitive pattern that appears in the printing.For this,firstly the module of the transformed image is calculated and then the logarithm is applied to obtain a higher contrast transformed image.A histogram of the resulting image is calculated to select a percentage of the existing frequencies.In the tests,this percentage oscillated around15%.In other words,for the filtering,the algorithm will choose to eliminate a percent-age of the frequencies that dynamically exceed a threshold obtained from the windows analysis.Once the threshold has been obtained,these frequencies are eliminated and the inverse transform is applied again to obtain thefiltered im-age.3.2.3Post-processingAfter thefiltered image is obtained,the post-processed one is calculated,which includes the detection of possible de-fects and if this has been repeated.For the detection of possible defects,a similar method to the FFTfiltering is used.The image histogram is calculated and is analyzed to remain with a certain percentage of pixels.Then the image is thresholded,the isolated background noise pixels appear, and only one pixel sized items are removed,since the real defects always have a greater size.When we have the real defects that appear in the im-age,itfinds out if they are significant defects for the com-pany or not.In this case,a defect is only significant,if itrepeats during all rapports,since these take place by the ob-struction of some of the points of the cylinder injector.If the defect is not repeated,the obstruction has disappeared by itself and it does not make sense to stop production to correct it.For that reason,when we detect a defect,it is analyzed and only if this has already appeared in previous rapport is the defect marked to be significant.If the defect is new,it is stored so that its possible repetition can be ver-ified.If the defect doesn’t appear in the following rapports, it will be eliminated because it is not considered significant.3.2.4ImplementationThe inspection system has two different parts:The me-chanical structure and the hardware that are responsible for the correct acquisition of the images of the fabric,and the software application that does all the processing,analysis and possible actions to be taken.The system has to analyze all the fabric,and when some relevant defects are found(as has been commented previously,in the case of the print-ing,the relevant defects are those that take place in each rapport and not in a sporadic form),the machine can make the decision to stop production so that the worker can clean the stamping roller.When the machine starts to analyze the fabric,the program starts a thread to do the acquisition of the images and whenever a new image is acquired,another thread is started to process the frame.This thread analy-ses the image using the technique explained in the previ-ous section and gives back a report with the possible errors found on it.Another problem found,is due to the use of a linear camera that causes due to optical and luminance rea-sons,that the ends of the image always appear darker than the central part.Nevertheless it is a smaller problem since the change of intensity for this reason isfixed and known. To solve this problem,a correction vector is captured from a white image and is stored and used later to correct all the acquisitions.The designed prototype can work as well in colour images,as in black and white images,although the optimal results are obtained normally using colour images. Although the captured image is a colour one,the analysis is made with a black and white transformed image.For the conversion,an algorithm of maximum contrast is used that gives very good results even with the mixtures of difficult colours such as blue or black base with black design.Infig-ure2,there is a dark green color design,with the correction and contrast stretch at the bottom.Once the corrected im-age is made,the different windows are calculated to do the FFT and,thenfilter it,as has been explained in the previous section.Infigure3,thefiltered result of the whole image can be observed.In the image,a black point is appraised that exactly represents the defect that exists in the design and that thefiltration heightens when eliminating the rest of the repetitive design of thepattern.Figure2.ImageCorrectionFigure3.FTT Filtered Image4Results and ConclusionsIn thefigure4,the defects detected by the system are shown.There are a lot of repetitive patterns that have been tested correctly by thefiltration algorithm.As can be seen in the images,the proposed algorithm works very well with all the tested designs even with very slight imperfections and strange asymmetric designs.It also has been tested with a wide colour spectra,including colours difficult to the human eye such as dark blue adhesive with dark blue weave,and black on black,and the results are good.5References[1]R.Meier,J.Uhlmann,R.Leuenberger.El sistema auto-matico de inspeccion para tejidos.Revista de la Industria Textil,Zellweger Uster,375,2000.[2]Shaw-jyh Shin,I-Shou Tsai,Po-Dong Lee.Feng Chia.Automatic faults detection and recognition for static plain fabrics.I nt.Journal of Clothing Science and Tech-nology.8(1/2),56-65.[3]J.G Campbell,Fionn Murtagh.Automatic Visual Inspection of Woven Textiles Using a Two-stage Defect Detector.Optical Engineering,37(9),1998,2536-2542.[4]J.G.Campbell,A.A.Hashim,T.Martin,McGin-nity Thomas,F.Lunney.Flaw Detection in woven tex-tiles by neural network.N eural Networks Conference, Maynooth,1995,92–99.[5]J.G.Campbell,A.A.Hashim,F.D.Murtagh.Flaw Detection in Woven Textiles using Space-dependet Fourier Transform.Irish Signals and Systems Conference,London-derry,N.Ireland,1997,500-506.[6]J.G.Campbell, C.Fraley, F.Murtagh, A.E. Raftery.Linear Flaw Detection in woven textiles usingFigure4.Examples of Defect DetectionModel-based Clustering.Irish Machine Vision and Image Processing Conference,Londonderry,N.Ireland,1997, 241-252.[7]F.S.Cohen,Zhigang Fan,Stephane Attali.Auto-mated Inspection of Textile Fabrics Using Textural Models. IEEE Transactions on Pattern Analysis and Machine Intel-ligence,13(8),1991,803-808.[8]H.Balakrishnan,S.Venkataraman,S.Jayaraman. FDICS.A Vision-based System for the Identification and Classification of Fabric Defects.Journal of Textile Insti-tute,1998.[9]Lois M.Hoffer,Franco Francini,B.Tiribilli,G. Longobardi.Neural networks for the optical recognition of defects in cloth.Optical Enginers,35(11),1996,3183-3190.[10]J.Lewis Dorroty,G.Vachtsevanos,Warren Jasper.Real-Time Fabric Defect Detection and Control in Wearing Processes.National Textile Center Annual Report, Georgia Institute of Technology,North Carolina State Uni-versity,1995,143-152.[11]lan,J.Escofet,J.Pladellorens,R. Navarro.Recognition and Inspection of Textile webs us-ing Fourier Analysis and Gabor Filters.VII National Symposium on Pattern Recognition and Image Analysis, Barcelona,Spain,1997,299-304.。