Textile Research Journal

http://trj.sagepub.com Computer Vision-Aided Fabric Inspection System for On-Circular Knitting Machine
R. Ghazi Saeidi, M. Latifi, S. Shaikhzadeh Najar and A. Ghazi Saeidi Textile Research Journal 2005; 75; 492 DOI: 10.1177/0040517505053874 The online version of this article can be found at: http://trj.sagepub.com/cgi/content/abstract/75/6/492

Published by:
http://www.sagepublications.com

Additional services and information for Textile Research Journal can be found at: Email Alerts: http://trj.sagepub.com/cgi/alerts Subscriptions: http://trj.sagepub.com/subscriptions Reprints: http://www.sagepub.com/journalsReprints.nav Permissions: http://www.sagepub.co.uk/journalsPermissions.nav Citations http://trj.sagepub.com/cgi/content/refs/75/6/492

Downloaded from http://trj.sagepub.com by Ershad Khan on October 21, 2008

492

TEXTILE RESEARCH JOURNAL

Computer Vision-Aided Fabric Inspection System for On-Circular Knitting Machine

R. GHAZI SAEIDI
Textile Engineering Group, Azad University, South Tehran Branch, Iran.

M. LATIFI1

AND

S. SHAIKHZADEH NAJAR

Textile Engineering Department, Amirkabir University of Technology, Tehran, Iran.

A. GHAZI SAEIDI
Electronic Engineering Department, Mazandaran University, Babol, Iran. ABSTRACT This paper describes a computer vision-based fabric inspection system implemented on a circular knitting machine to inspect the fabric under construction. The study consisted of two parts. In the rst part, detection of defects in knitted fabric was performed and the performance of three different spectral methods, namely the discrete Fourier transform, the wavelet and the Gabor transforms were evaluated off-line. In the second part, knitted fabric defect-detection and classication was implemented on-line. The captured images were subjected to a defect-detection algorithm, which was based on the concepts of the Gabor wavelet transform, and a neural network (as a classier). An operator encountering defects also evaluated the performance of the system. The fabric images were broadly classied into seven main categories as well as seven combined defects. The results of the designed system were compared with those of human vision.

Circular knitting is one of the easiest and fastest ways (20 million stitches per minute) to produce cloth and textile pieces such as garments, socks, and gloves. The fabric roll is removed from the large diameter circular knitting machine, and then sent to an inspection frame. If inspection could be done on the machine, the need for 100% manual inspection would be eliminated [7]. The existing defect-detection techniques can be classied into three different categories: statistical, spectral, and model-based. The defect-detection approach by Zhang and Bresee [21] is based on rst-order statistics such as mean and standard deviation. The fabric image is divided into sub-blocks with the use of information obtained by auto-correlation. The use of a gray-level cooccurrence matrix of the image is based on second-order statistics [2]. However, these statistical techniques are not useful for the detection of those textured defects whose statistical features, namely the rst- and secondorder moments, are signicantly close to that of defectfree textured regions [13]. A high level of quality assur-

1 To whom lati@aut.ac.ir

correspondence

should

be

addressed:

e-mail:

ance requires identication of such defects, and therefore techniques based on spectral features have been investigated in the literature. Textured materials, such as woven and knitted fabrics, possess strong periodicity due to the repetition of the basic weaving pattern. Therefore spectral techniques using a discrete Fourier transform [4, 19], optical Fourier transform [17], and windowed Fourier transform [3] have been used to detect woven fabric defects. Escofet et al. [8] have used the angular correlation of the Fourier spectra to evaluate fabric web resistance to abrasion. Sari-Saraf and Goddard [18], and Chan and Pang [4] have used a Fourier transform to detect fabric defects. Ravandi and Toriumi [16] have also used Fourier transform analysis to measure fabric appearance. Escofet et al. [9] have used a bank of multi-scale and multi-orientation Gabor lters for the detection of local fabric defects. Kumar and Pang [12] have demonstrated another approach to fabric defect detection using real Gabor functions. Jasper et al. [10] have shown that the texture information can be adapted into wavelet bases. Such an adapted wavelet bases offer high sensitivity to the abrupt changes in the surface texture caused by defects, enabling their detection. Several works have also demon 2005 Sage Publications www.sagepublications.com

Textile Res. J. 75(6), 492 497 (2005) DOI: 10.1177/0040517505053874

Downloaded from http://trj.sagepub.com by Ershad Khan on October 21, 2008

JUNE 2005 strated various approaches to the detection of surface defects using wavelet basis functions [15, 11, 14]. The Markov random eld model is a model base for describing local statistical dependence of the image [5, 6]. The present work focuses on defect-detection and classication on an on-circular knitting machine during processing using spectral methods including the discrete Fourier transform, wavelet and Gabor.

493 the discrete Fourier transform, the wavelet and Gabor transforms for fabric fault detection. In order to control the image acquisition process, a program using LabView software was developed. The program was responsible for on-line capturing and storing of digital images received from the CMOS camera according to the microcontroller commands. Defect-detection and classication was performed off-line using a program written with MATLAB software. The knitting machine was run with different speeds (2/3 to 17/rpm) and different fabric faults including vertical stripe, vertical soil stripe, horizontal stripe, horizontal soil stripe, crack or hole, nep or slub were created. Images of these fabric samples were acquired under front lighting. Figure 1 shows typical fabric defects considered in this experiment. The dimension of each individual image was 3.7 3.4 cm (176 144 pixels), with 256 gray levels. A total of 2315 fabric images were captured and then the fabric faults were detected off-line. As a front-lighting method was used, there were no very noticeable differences between the vertical faults resulting from various causes. For this reason, these faults were classied in the vertical stripe category. A similar situation was found between the loop length variation and double yarn fault so that they were classied in the horizontal faults group. In this experiment, 20 images corresponding to six different categories of fabric defect and 20 images of defect-free fabric were used. In general, 50% of the images of each category were employed for training and remaining images were used for testing. THE FOURIER TRANSFORM METHOD In this section, a method that was very similar to those of Ribolzi et al. [17] and Tsai and Hu [20] was used for detecting fabric defects. To classify fabric faults, a three-layer perceptron neural network with a feed-forward, back- propagation al-

Hardware Description
KNITTING MACHINE SPECIFICATIONS A single jersey mini-jacquard circular knitting machine with a 30 inch cylinder, gauge 24 and 72 feeders was used. On this machine, a plain knitted fabric can be produced using viscose/polyester yarn (30 Ne). Due to some operational limitations for mounting the camera and in order to capture the full fabric surface image, only 36 yarn feeders were utilized. IMAGE ACQUISITION SYSTEM The system consisted of a 640 320 element CMOS (Complementary Metal Oxide Semiconductor) camera, which was synchronized to the moving fabric by means of an incremental encoder, a microcontroller board, and a personal computer-based AMD Athlon XP 1800 processor. The custom-made component in this system was the microcontroller board, which was used to extract the forward movement of the fabric and to enable accurate image capture by the camera. These components were used to acquire images of the fabric under construction and to store them on the computer.

Experiments
LABORATORY DEVELOPED EXPERIMENT The aim of this test was to evaluate and compare the performance of three different spectral methods, namely

FIGURE 1. Images of defective fabrics.

Downloaded from http://trj.sagepub.com by Ershad Khan on October 21, 2008

494 gorithm (three neurons in the rst, 10 neurons in the second and seven neurons in the third layer, with a learning rate of 0.025 and 3750 epoch) was applied. The results of the Fourier transform method are presented in Table I. THE WAVELET TRANSFORM METHOD A wavelet transform based on Daubechies [1], second class (equation (1)), was employed. The experimental results indicated that an appropriate classication could be achieved using the third level of the transform. G 4 2 1 1 3 ej3 3 3ej2 3 3ej 1 3 , (1)

TEXTILE RESEARCH JOURNAL

TABLE II. Parameters value of Gabor wavelet transform. Parameters Number of wavelet scales Number of lter orientations Wavelength of smallest scale lter Scaling factor between successive lters Ratio of the standard deviation of the Gaussian describing the Gabor lters transfer function in the frequency domain to the lter center frequency Ratio of angular interval between lter orientations and the standard deviation of the angular Gaussian function used to construct lters in the frequency plane Orientation Scale Value 4 6 3 4 0.35 1.9 1,3,4,6 2

SUMMARY

OF

LABORATORY EXPERIMENTAL RESULTS

where, G() is Daubechies function and is variable. To classify imperfect fabrics, a three-layer perceptron neural network with a feed-forward, back-propagation algorithm (30 neurons in the rst, eight neurons in the second and seven neurons in the third layer with a learning rate of 0.275 and 2250 epoch) was utilized. The detailed image was used to calculate the mean and variance of its pixel intensities in order to train the neural network. The results of the wavelet transform method are presented below in Table I. THE GABOR TRANSFORM METHOD In the last part of the laboratory experiments, a Gabor wavelet transform based on the work of Escofet et al. [9], was designed with the parameter values as shown in Table II. A three-layer perceptron neural network with a feedforward, back-propagation algorithm (15 neurons in the rst, eight neurons in the second and seven neurons in the third layer with a learning rate of 0.475 and 7350 epoch) was used as a classier. The results of the Gabor transform method are presented in Table I.

From the results presented in Table I, it can be concluded that the Gabor transform method with a success rate of 78.4% had the highest efciency value of the three methods. Therefore, the Gabor transform method together with a three-layer perceptron neural network and a feed-forward, back-propagation algorithm was selected to continue the experiment in an industrial system. ON-LINE IMPLEMENTATION
ON AN INDUSTRIAL

SYSTEM

In this part of the study the whole process of image capturing, defect detection and classication was performed on-line. A schematic diagram of the experimental set-up is shown in Figure 2. In order to capture, store and process the images on-line, at a rate of 13 frames per second (fps), two computers with similar specications were used. The rst computer received and detected images using two linked programs (Lab-View for image capturing and MATLAB for image detection). The second computer was utilized for fabric defect classication and grading using two linked programs (Lab-View for computer communication and fabric grading and MATLAB for defect classications). In this experiment, 50 images corresponding to six different individual categories of fabric defects and 50

TABLE I. The success rate of fabric defect detection corresponding to the different spectral methods. Fourier transform Class type of fabric images Vertical stripe Horizontal stripe Vertical soil stripe Horizontal soil stripe Nep/slub Hole Defect free Overall success rate (%) No. of detected images 1 5 1 1 3 0 0 15.71 Wavelet transform No. of detected images 7 5 5 10 5 4 0 52.3 Gabor transform No. of detected images 10 10 2 10 9 8 6 78.4

Downloaded from http://trj.sagepub.com by Ershad Khan on October 21, 2008

JUNE 2005

495

FIGURE 2. Experimental set-up for industrial system.

images of defect-free fabric selected from the 2315 images were applied for neural network training. The knitting machine was run at speeds of 2/3 to 10 rpm and 8135 images including different types of defects and perfect fabric were captured and analyzed using the proposed program. To precisely investigate and maintain the efciency of this program, all 8135 images were visually inspected and nally the images corresponding to the defect types were categorized and classied. The results of the program were compared with those of human vision inspection and then the success rate of the system was determined (Table III).

Results and Discussion

The experimental results presented in Table I demonstrate that the application of a Gabor transform method

designed for the present work were highly promising in terms of the identication of defects in knitted fabric, with an overall success rate of 78.4%. In the case of vertical and horizontal stripes, horizontal soil stripes and nep/slub defects, the success rate was more than 90%. However, the vertical soil stripe was detected at a lower rate (20%) because the intensity of this defect was close to that of the fabric surface (so it is hard to distinguish it from a defect-free fabric). In the case of defect-free samples, the success rate was 60%, presumably due to the low number of samples used for training the neural network. The results from the analysis of 8135 images using the industrial system for both defective and defect-free fabric are shown in Table III. They show an overall success rate of 96.57% with a location accuracy of 2 mm and a

TABLE III. Final experimental results for industrial system No. of images 163 127 114 120 1722 14 5791 19 3 6 14 28 4 10 8135 No. of true detected images 132 104 67 98 1651 9 5735 16 2 3 8 21 2 8 7856 No. of undetected images 1 0 45 12 10 3 0 2 0 3 2 1 1 1 163 No. of false detected images 30 23 2 10 61 2 56 1 1 0 4 6 1 1 116 Success rate (percentage) 80.98 81.88 58.77 81.66 95.87 64.28 99.03 84.21 66.66 50 57.14 75 50 80 96.57

Class type of fabric image Vertical stripe Horizontal stripe Vertical soil stripe Horizontal soil stripe Nep/slub Hole Defect free Vertical and horizontal stripe Horizontal soil stripe and nep Vertical soil stripe and nep Vertical stripe and nep Horizontal stripe and hole Vertical stripe and hole Horizontal stripe and nep Total results

Downloaded from http://trj.sagepub.com by Ershad Khan on October 21, 2008

496 false alarm rate of 1.4%. It should be noted that the false alarm rate was computed as the total number of false detections divided by the total number of processed images. It can be seen that for the most common and the most serious defects, such as vertical and horizontal stripes, vertical and horizontal soil stripes, neps/slubs, horizontal-vertical stripes, horizontal stripes-holes and horizontal stripes-neps defects, the system performance was acceptable with the success rates of 75 to 95.87%. On the other hand, other defects such as vertical soil stripes, vertical soil stripes-neps, vertical stripes-neps, and horizontal stripes-holes are detected at a lower success rate of 50 to 60%. In the case of holes defects, they appeared at the edge of captured image and were miss-identied as other defect types. As the intensity of vertical soil stripes was close to that of the defect-free fabric surface, they were detected at a low rate of 58.77%. This result in turn causes the success rate of the vertical soil stripes-neps defects to also be low (50%). Furthermore, as the position of the vertical stripe and nep defects in some images were very close, these defects are mixed together leading to a low success rate of 57.14%. For the case of the vertical stripe-hole defect, the number of samples was too low and needs further investigation. Typical images of fabric and detected defects are shown in Figure 3.

TEXTILE RESEARCH JOURNAL

Conclusions
A computer vision-aided fabric inspection system for an on-circular knitting machine has been described. Two experimental approaches, a laboratory-developed experiment and an industrial system were investigated. In the case of the laboratory experiment, in order to detect fabric faults, three different spectral methods (discrete Fourier transform, the wavelet and the Gabor transforms) were applied to fabric images. It was found that the Gabor transform method with a success rate of 78.4% has the highest efciency value among three methods. In the case of the industrial system, fabric images including different types of defects and also defect free were captured and analyzed using the proposed program. The captured images were then detected by applying the Gabor transform and classied into 14 groups utilizing a neural network. To precisely investigate the efciency of the system, all images were inspected under human vision and nally were classied corresponding to defect types. The results of the developed system were compared with those from the visual categorization and then the success rate of the designed machine vision system was calculated. The overall success rate of the proposed approach was found to be 96.57% with a location accuracy of 2 mm and a false alarm rate of 1.4%. This paper has presented a new approach for online detection of knitted fabric defects using the Gabor trans-

FIGURE 3. Images of fabric and detected defects.

Downloaded from http://trj.sagepub.com by Ershad Khan on October 21, 2008

JUNE 2005 form. Since this research is limited by the speed of the knitting machine, further studies are required to inspect the fabric defects in higher speed, circular knitting machines. ACKNOWLEDGMENTS Thanks are due to Ministry of Industries and Mines of Iran and the Center of Excellence in Textile Engineering of Amirkabir Uni. of Technology for support of this work.

497
J., Detection of Local Defects in Textiles Webs using Gabor Filters, Opt. Eng. 37, 22972307 (1998). Jasper, W. J., Garnier, S. J., and Potapalli, H., Texture Characterization and Defect Detection using Adaptive Wavelets, Opt. Eng. 35, 3140 3149 (1996). Kim, S., Lee, M. H., and Woo, K. B., Wavelet Analysis to Fabric Defects Detection in Weaving Processes, in Proc. IEEE Int. Symp. Industrial Electronics 3, 1406 1409 (1999). Kumar, A., and Pang, G., Fabric Defect Segmentation using Multichannel Blob Detectors, Opt. Eng. 39, 3176 3190 (2000). Kumar, A., and Pang, G., Identication of Surface Defects in Textured Materials using Wavelet Packets, Proc. 36th IEEE/IAS Annual Meeting, Illinois, September, 2001. Lambert, G., and Bock, F., Wavelet Methods for Texture Defect Detection, Proc. IEEE Intl. Conf. Image Processing. 3, 201204, (1997). Mufti, M., Fault Detection and Identication using Fuzzy Wavelets, PhD Thesis, Dept. of Electrical and Computer Engineering, Georgia Institute of Technology, 1995. Ravandi, S. A. H., and Toriumi, K., Fourier Transform Analysis of Plain Weave Fabric Appearance, Textile Res. J. 65, 676 683 (1995). Ribolzi, S., Merckle, J., and Gresser, J., Real-Time Fault Detection on Textiles using Opto- Electronic Processing, Textile Res. J. 63, 61 67 (1993). Sari-Sarraf, H., and Goddard, J. S., Vision Systems for On-loom Fabric Inspection, IEEE Trans. Ind. Appl. 35, 12521259 (1999). Tsai, D. M., and Heish, C. Y., Automated Surface Inspection for Directional Textures, Image and Vision Comput. 18, 49 62 (1999). Tsai, I. S., and Hu, M. C., Automated Inspection of Fabric Defects using an Articial Neural Network Technique, Textile Res. J. 66, 474 482 (1996). Zhang, Y. F., and Bresee, R. R., Fabric Defect Detection and Classication using Image Analysis. Textile Res. J. 65, 19 (1995).

10.

11.

12.

Literature Cited
1. Abbate, A., Decusatis, C. M., and Das, P. K., Wavelets and Subbands Fundamentals and Applications, Birkhauser, Boston, 2002. 2. Amett, A. L., Texture Defect Detection Using Subband Domain, in IEEE Southwest Symposium on Image Analysis and Interpretation, 205210 (1998). 3. Campbell, J. G., and Murtagh, F., Automatic Visual Inspection of Woven Textiles using a Two-stage Defect Detector, Opt. Eng. 37, 2536 2542 (1988). 4. Chan, C. H., and Pang, G., Fabric Defect Detection by Fourier Analysis, IEEE Trans. Ind. Appl. 36, 12671276 (2000). 5. Chellappa, R., and Chatterjee, S., Classication of Textures Using Gaussian Markov Random Fields, IEEE Trans. On Acoust., Speech, Signal Processing. 33, 959 963 (1985). 6. Cross, G., and Jain, A., Markov Random Field Texture Models, IEEE PAMI. 5, No. 1, (1983). 7. Dorrity, J. L., Vachtsevanos, G., and Jasper, W., Real-time Fabric Defect Detection and Control in Weaving Processes, National Textile Center Annual Report, 113122 November, 1996. 8. Escofet, J., Millan, M. S., Abril, H., and Torrecilla, E., Inspection of Fabric Resistance to Abrasion by Fourier Analysis, Proc. SPIE 3490, 207210 (1998). 9. Escofet, J., Navarro, R., Millan, M. S., and Pladelloreans,

13.

14.

15.

16.

17.

18.

19.

20.

21.

Downloaded from http://trj.sagepub.com by Ershad Khan on October 21, 2008