Journal of Textile Institute, 88, 488-500, 1997.
INSTRUMENTAL EVALUATION OF FABRIC PILLING Bugao Xu Department of Human Ecology, University of Texas, Austin, TX 78712
ABSTRACT Image analysis has been widely accepted as an objective method for evaluating fabric appearance. This paper presents the development of an image analysis system that aims at characterizing and rating fabric pilling appearance. The procedures and examples of using the Fast Fourier Transform (FFT) and other techniques to correct image defects, such as nonuniform background and low contrast, are shown. A novel approach that splits a fabric image into periodic and non-periodic structures is explained. A template-matching technique used for extracting pills from the non-periodic image is discussed. Pill properties are characterized by density, size and contrast, and the measurements on the ASTM photographic standards are used to build the empirical grading equations of fabric pilling. The pilling of textile fabrics refers to an appearance caused by bunches or balls of tangled fibers held to the surface. This unpleasant appearance can seriously compromise the fabrics’ acceptability for apparel. Pills are developed on a fabric surface in four main stages: fuzz formation, entanglement, growth, and wear-off [3]. In normal wear, a piece of a garment may take a long time to be pilled. To expedite the pilling evaluation, a number of testing machines have been designed to simulate the pilling that occurrs in normal wear. Fabrics are forced to form typical pills by tumbling, brushing, or rubbing specimens with abrasive materials in machines, and then are compared with visual standards, which may be actual fabrics or photographs of fabrics, to determine the degree of pilling on a scale ranging from 5 (no pilling) to 1 (very severe pilling) [1,2]. In the ASTM D3511 and D3512 pilling resistance test methods, an observer is guided to assess the pilling appearance of a tested specimen based on a combined impression of the density and size of pills, and degree of color contrast around pilled areas. Counting the pills and weighting their number with respect to their size and contrast as a combined measure of the pilling appearance is not recommended, because this requires excessive time if done manually [1]. A frequent complaint about the visual evaluation method is its 1
Journal of Textile Institute, 88, 488-500, 1997. inconsistency and inaccuracy of the rating results. More reliable and objective methods for pilling evaluation are desirable for the textile industry. Computer vision technology provides one of the best solutions for the objective evaluation of pilling. Researchers in various institutions have been exploring image analysis techniques effective for pill identification and characterization [6, 8, 9]. A typical set-up of an image analysis system for pilling evaluation includes a CCD camera, frame grabber, computer and the analysis software. The suitable hardware is generally available on the market, but the analysis software often requires a great deal of customization or new development. A pilling evaluation program needs to perform at least two procedures: pill identification and feature measurement. The simple computation algorithms involved in pill identification in a solid-color fabric follows a common principle that pills are brighter than their surrounding areas, and therefore can be differentiated by selecting a proper threshold. As pointed out in [8], it may be very difficult to identify pills in the image of a patterned fabric captured with a CCD camera since pills appear in different intensities over differently-colored regions. Rangulam et. al. used a laser scanning mechanism to acquire fabric images to avoid the difficulty in identifying pills on a patterned fabric. The laser scanning, however, is a much slower process than the camera capturing, since it needs an x-y stage to mechanically transport the sample. Ideally, an image analysis system developed for pill evaluation should be consistent with current visual standards, applicable to samples that have different colors or patterns, independent of the system operators, and reasonably fast. To achieve these goals, the following issues still need to be addressed. How can pills be correctly identified and extracted? What features of pills need to be measured to characterize pilling appearance? What relationships are among the features? If a CCD camera or scanner is used, how does the system deal with patterned fabrics? 2
Journal of Textile Institute, 88, 488-500, 1997. The present research attempts to seek solutions for these problems. Since removing colored patterns in a patterned image is still part of the on-going research, discussions regarding this topic will be held for a later report.
IMAGE SYSTEM SYSTEM The image analysis system used in the research follows the basic set-up described in previous publications [12], except for the following changes. To accommdate for differences in fabric color, an auto-iris lens is used so that a consistent brightness can be maintained over dark and bright samples. A 3-chip color CCD camera is used to provide accurate color information for pattern-removing in the continuing study. Since pills observed in worn garments vary appreciably in size and appearance, the system should be capable of capturing and analyzing multi-frame images of the sample at various locations to generate more reliable statistical data. To avoid the human interference, a stage was designed to automatically transport the sample under the camera (Figure 1). The stage, driven by a DC motor, moves in one direction and has a travel of 12 inches. Two limit switches are mounted at each end of travel to prevent false operations. The number of positions and the interval between two positions can be set from the computer interface (Figure 2). The computer checks and adjusts the input interval to warrant the total moving distance not to exceed the maximum travel. The stage starts to move from one end, and stops at each position to let the camera capture a still image. The pause time is adjustable using the software. When the stage moves to the next position (the time needed depends on the interval distance), the computer takes the chance to conduct certain image-processing and measurement tasks. All these movements are controlled by the computer and a specially-
3
Journal of Textile Institute, 88, 488-500, 1997.
Figure 1 Mechanical Stage and 3-CCD Camera
designed circuit that is connected to the computer parallel port. This simplified stage greatly cuts the price of commercially available mechanical stages, while perfectly satisfying the need for positioning the fabric sample.
Figure 2 Stage Control
4
Journal of Textile Institute, 88, 488-500, 1997.
IMAGE ENHANCEMENT Since the development of pills may be accompanied by other surface phenomena such as loss of cover, color change, or the development of fuzz [1], the image of a tested fabric often contains non-uniform background, varying contrast, and other defects. It is necessary to correct or reduce the image defects to facilitate pill identification. Several examples of the functions that are effective in enhancing images are briefly discussed here. More details can be found from the author’s previous publications [11] and image processing references [5, 10]. A fabric may result in a low-contrast image under a normal image capturing condition in which yarn structures, pills and other details are barely visible. For a low-contrast image, a contrast mapping technique can be used to maximize the contrast of the image [5, 11]. The reason for the low contrast in an image is that the image histogram takes advantage of only the narrow band of the gray-scale range that a computer system provides. A modern PC computer can easily display 256 gray levels simultaneously (255 for white and 0 for black). One can
a
b
original enhanced Figure 3 Contrast Stretching perform the following mapping to make optimal use of the available gray-scales when displaying 5
Journal of Textile Institute, 88, 488-500, 1997. the image. The darkest pixels in the original image are forced to be black, the brightest pixels to be white, and an intermediate gray level to be a value which is linearly interpolated between black and white. This mapping will not change the number of gray levels and the relative frequency of each level, but will stretch the histogram to cover the entire gray-scale range (see the histogram of the enhanced image in Figure 3). Having been processed by this method, the enhanced image clearly exhibits the fabric structure and pills. Another common problem in processing a pilled-fabric image is the non-uniformity of the overall intensity. A gradual variation in intensity may exist in the image due to non-uniform illumination. For example, the center part of image b in Figure 3 appears brighter than other regions. This slow change can be considered as the background of the image, since its variation does not reflect the interesting texture of the image. This defect can be minimized by fitting a background function into the image [10, 11]. It takes three steps to conduct the background leveling. First, a number of pixels are sampled throughout the image. When deciding how many pixels should be sampled, two conflicting factors, fitting accuracy and computation time, should a
b
background
leveled Figure 4 Background leveling
6
Journal of Textile Institute, 88, 488-500, 1997. be balanced. It seems appropriate to sample 1/5 to 1/10 of the total pixels of an image. Then, a two-dimensional function that approximates the background is determined by using the leastsquare fitting technique [11]. A second-order polynomial function is one of the most common functions used for the background approximation. Finally, the background function is subtracted from the image, and a new image that does not contain the background variation is generated. Figure 4 shows the background function of image b in Figure 3 and the leveled image. Some pilled fabrics may contain complex color variation (shading) that the leveling technique described above is unable to correct (image a in Figure 5). Fourier filtering is an effective way to eliminate this type of defect [10, 14]. In general, shading is a more irregular, slower intensity change than the intensity alteration of the fabric structure, and therefore can be classified as low-frequency components. A power spectrum showing the contributions of all frequency terms to the image can be obtained by performing the Fourier transform. Frequency terms in the spectrum are independent and separable. Therefore, one can delete the frequency terms in a particular region which result from undesirable signals in the original image. The modified power spectrum is used to reconstruct the image through the inverse Fourier transform. Frequency terms corresponding to shading should be concentrated in a region near the origin of the spectrum (low frequencies). A circle centered at the origin may be used to select these frequency terms. If the terms inside the circle are excluded when reconstructing the image, shading will not appear in the reconstructed image. Obviously, the size of the circle is crucial. For the purpose of reducing the degree of shading, the circle should not enclose any prominent peaks because they represent the structures of the image. Figure 6 shows one example of applying Fourier filtering to suppress shading. Image a is one of the ASTM pill photographs, image b is its power spectrum, image c displays the modified power spectrum, and image d 7
Journal of Textile Institute, 88, 488-500, 1997. shows the reconstructed image. It is worthwhile to reiterate here that the Fourier transform and its inverse transform can be greatly expedited by the application of a fast Fourier transform algorithm (FFT).
a
b
c
d
Figure 5 Image Shading Corrected by Fourier Filtering, a: original; b: power spectrum; c: modified spectrum; d: corrected
PILL EXTRACTION Pills often appear in comparable brightness and size to those of floating yarns. It is extremely difficult to discern pills from yarn points simply by imposing the intensity and/or size thresholds. They, however, possess totally different spatial structures from which a new differentiation clue may be derived. Pills are localized minor disturbances [3] randomly distributed on the surface, while yarn floating points, which are part of the fabric structure, appear to be periodical and associated only with the pattern of interlacing (weave) or 8
Journal of Textile Institute, 88, 488-500, 1997. interlooping (knit). The periodic structure of a fabric will result in prominent peaks in its power specturm, and non-periodic components such as pills will generate frequency terms spreading in the background of the spectrum. Images a in Figure 6 shows the power spectrum of Images b in Figure 4 obtained through the 2D FFT (see more detailed explanations about spectrum in a previous paper [14]). Since it is much more convenient and precise to locate peaks in the spectrum than to find the periodic structure in the original image, the FFT technique can play an important role in extracting pills.
a
b
c
d
Figure 6 Pill Extraction a: power spectrum, b: peak areas; c: non-periodic structure; d: periodic A region-growing method is then employed to locate peaks in the power spectrum. Peaks are small, bright regions that can be detected by using a power threshold. The threshold usually needs to be set to a reasonably high value, e.g., the mean power plus three times the standard 9
Journal of Textile Institute, 88, 488-500, 1997. deviation, to prevent noisy areas from being detected. The pixels whose powers are above the threshold are considered as the core parts of peaks. Each peak expands by merging with its neighboring pixels that are above half of the highest power in this peak region (see the black marks in image b of Figure 6). The spectrum is now divided into the non-peak portion and the peak portion, which can be used in the inverse FFT to reconstruct two images, respectively. The image reconstructed through the non-peak spectrum presents the non-periodic structure including pills (image c), and that from the peak spectrum shows the fabric weaving or knitting pattern (image d). The non-periodic image, of course, contains all noise elements, which differ significantly in size and shape. Since most pills appear to be solid circular regions, a templatematching technique is suitable to the further detection of pills in the non-periodic image. A template is a pictorial representation of a known feature. For pill detection, the template is designed to be a small square which contains a centered white circle surrounded by black pixels. Note that the size of the template should be equivalent to the size of the repeating units in the fabric, which can also be determined by using the FFT techniques [14]. Template matching is the process of moving the template over the entire image, and calculating the similarity between the template and the covered window on the image. The normalized correlation is one commonly used measure of similarity to determine match. If the image and the template are denoted as f(x,y) and t(x,y), the correlation coefficient at point (m,n) is given by [10]
r ( m, n) =
∑ ∑ [ f ( x, y) − f xy ][t ( x − m, y − n) − t ] x y
2 2 ∑ ∑ [ f ( x, y) − f xy ] [t ( x − m, y − n) − t ] x y
10
Journal of Textile Institute, 88, 488-500, 1997. where, f xy is the average intensity of image pixels within the window that is translated across the entire image, and t is the average intensity of the template. The double summations are carried out over the moving template and the covered window. This equation can be further simplified to reduce redundant calculations in the program as follows:
r ( m, n) =
M ∑ ∑ [ f ( x, y) t ( x − m, y − n)] − f xy t x y
2
[ M ∑ ∑ f 2 ( x, y) − f xy ][ M ∑ ∑ t 2 ( x − m, y − n) − t 2] x y
x y
here, M is the number of pixels in the template. After the template matching, a new image, named as a matching map, can be generated from r(m,n) to visually display locations at which the template best fits the image (image a in Figure 7). The bright areas in the matching map indicate pill locations, and can be readily segmented using a global threshold because the map is free from the fabric patterns and shading (image b in Figure 7). Tiny particles in the thresholded image need to be removed by running the morphological opening operation [4], since they correspond to fuzz (image b in Figure 7).
b
a
Figure 7 Template matching (a) and Thresholding (b)
11
Journal of Textile Institute, 88, 488-500, 1997.
PILL CHARACTERIZATION In the visual evaluation, observers intend to rate the pilling appearance of a fabric by comparing pill properties such as density, size and height, to those of the visual standards. After pill extraction, the computer can accurately measure these properties. DENSITY Pill density is the first impression that an observer probably will get when examining a pilled sample. The density is often estimated by the number of pills in a unit area. This definition is accurate only if pills are randomly or uniformly distributed over the area selected for counting pills [7, 13]. When clumping occurs the result will substantially vary with the area. A more rational estimator of pill density can be constructed based on the distances of pills to their nearest neighbors. The nearest distance of two pills is the length between the two centers. The center of a convex-shape object such as a pill can be given by its mass center, whose x-y coordinates are equal to the first-order moments [5]. Since the analyzed area is only a small portion of the entire sample, a random sampling procedure is needed for estimating the population density. Inside the mass-center image (Figure 8), the computer randomly generates a number of points (triangles). At each of these points, the computer searches a pill (dot) closest to the point, and measures the nearest distance ri. Inside a circle whose radius is ri, only one pill exists. Then, the computer searches the nearest pill to this found pill, and measures their nearest distance xi. Inside a circle whose radius is xi, two pills exist. After n random points are counted, the total areas of the two sets of circles are π ∑ (r i2) and π ∑ ( x i2) , respectively, and a
12
Journal of Textile Institute, 88, 488-500, 1997.
xi ri
Figure 8 The Nearest Distance population-density estimator D can be derived as follows [7]. D=
2n
π ∑ (r i2) ∑ ( x i2)
The denominator in the formula is the geometric average of the areas of the two sets of circles. This estimator is insensitive to the spatial pattern of the mass-center image [7].
SIZE The average size of pills is another important factor influencing pilling appearance. The computer can locate each pill, count the pixels in the pill, and calculate the following statistical data: mean, standard deviation, maximum, minimum and area percentage, which is equal to the ratio of the total area of pills to the image area. The size distribution curve can be calculated as well.
CONTRAST The contrast between a pill and its surrounding region reflects the height of the pill. In a gray-scale image, the contrast between two regions is measured by the difference in intensity. To locate surrounding regions of pills, pilled areas in image b of Figure 7 are first enlarged by
13
Journal of Textile Institute, 88, 488-500, 1997. several layers (morphological dilation [4]). The dilated (image a in Figure 9) and undilated (image b in Figure 7) images are then subjected to a logical operation: xor, in which two coincidental pixels in the two images result in a white pixel if both are identical, or a black pixel otherwise. Xoring generates a new image showing the surrounding regions of pills (image b in Figure 9). Based on the locations of pills and their surrounding regions, the computer can measure the intensities for these two regions in the original gray-scale image, and then calculate the contrast for each pill.
a
b
Figure 9 Dilated Pills (a) and Surrounding Regions (b)
PILLING EVALUATION In order to make the rating results generated by the pilling evaluation system consistent with the visual standards, the ASTM photographic pilling standards (Figure 10) were first analyzed using the system and the rating equations were built based on the measurements of pill properties of these photographs (Table I). Although the average size of pills has a decreasing trend when the pilling grade increases, there is no significant difference between grades 1 and 2. This is because pills will be worn off as their sizes increase to a certain level. Hence, the average pill size alone is not sufficient for rating pilled samples. The density and % area of pills show relatively coherent decreases with the pilling grade, though the relationships are non-linear 14
Journal of Textile Institute, 88, 488-500, 1997. TABLE I Pill Measurements of ASTM Photographic Standards Photo M 1 2 3 4 5
4.68 4.53 2.26 1.74 1.48
Area (mm2) % Area Count Density Contrast SD Max Min (per 10x10cm2) 3.28 13.56 1.12 3.19 20.3 1.85 1.29 5.51 1.02 0.86 4.24 1.02 0.55 2.68 1.02
3.57 2.27 0.60 0.29 0.15
97 64 34 21 13
116.66 77.98 27.15 23.15 14.53
36.56 52.36 43.75 45.08 37.27
* M: mean; SD: standard deviation; Max: maximum; Min: minimum
15
Gd 1.05 1.91 3.04 3.81 5.10
Gs
G
1.05 1.92 2.92 4.23 4.83
1.05 1.92 2.98 4.02 4.97
Journal of Textile Institute, 88, 488-500, 1997. 1
2
3
4
5
16
Journal of Textile Institute, 88, 488-500, 1997. Figure 10 ASTM Pilling Photographs (Figure 11). Due to the inconsistencies in the quality of the photographs, the contrast of pills does not show a constant trend with the changes in pilling grade, and therefore becomes unreliable if used as a rating factor. The empirical rating equations of pilling can probably be built on pill density and % area by using regression techniques. Since the non-linear regression (negative exponential) did not provide a good fit, the segment linear regression was conducted to generate lines that fit two separate data groups (Figure 11). The rating equation based on the density D measurements is:
5 3. 64 − 2. 22 * 10 D ( D > 27 per10 × 10 cm ) G =6 7. 28 − 0.15 D ( D ≤ 27 per10 × 10 cm ) 7 and, the one based on % area S measurements is: 3. 44 − 0. 67 S ( S > 0. 6) 5 G =6 5. 47 − 4. 25S ( S ≤ 0. 6) 7 −2
2
d
2
s
here, Gd and Gs are the two different grades. Gd and Gs are assigned to zero when their calculated values are negative. The final grade G for the pilling appearance may be given by the average of
Gd and Gs. If a calculated G passes the low limit (1) or the high limit (5), it will be set to the limit. Table I also shows the rating grades of the ASTM standards by the system. The rating error is under 0.1. 5
5
4
4
Gd 3
Gs 3
2
2
1
1 0
20
40
60
80
100
0
120
1
2
AREA (%)
D (10*10 cm2)
Figure 11 Pill Measurements for ASTM Pilling Standards 17
3
4
Journal of Textile Institute, 88, 488-500, 1997.
RESULTS Four washed socks (s1~s4) and four naturally worn garmets (g1~g2) were tested using the system. The samples represent various degrees of pilling, and have different colors and structures (Figures 12 & 13). Table II shows the measured results. Besides the detailed information about pill properties, the computer also gave a rating of pilling appearance, which highly agreed with the visual impression. In some cases (e.g., s3 and g3), Gds are significantly different from Gss. This is because some fabrics tend to generate a larger number of pills with smaller sizes than others. G, combining both factors, should be more realistic. g2 has relatively few pills, which explains the high ranking by density and % area, but its pills are very intense. If the contrast is taken into account, g2 should have a lower pilling grade. Unfortunately, the current rating equations do not reflect the effects of the contrast.
s1
s2
s3
s4
Figure 12 Washed Socks (s1~s4), top row: original images; bottom row: pills
18
Journal of Textile Institute, 88, 488-500, 1997.
g1
g2
g3
g4
Figure 13 Worn Garments (g1~g4), top row: original images; bottom row: pills TABLE II Pill Measurements of Fabric Samples Sample M
s1 s2 s3 s4 g1 g2 g3 g4
1.50 1.78 2.39 3.87 1.87 1.31 1.31 1.47
Area (mm2) SD Max Min
0.39 0.96 1.29 2.24 0.62 0.55 0.54 0.65
2.32 5.00 8.54 9.76 2.93 2.56 2.68 2.93
0.98 0.85 0.97 1.46 0.73 0.85 0.72 0.73
% Area Count
0.17 0.87 2.10 4.65 0.63 0.13 1.54 0.87
9 39 70 96 27 8 94 47
Density Contrast 2 (per 10x10cm )
17.65 79.21 147.03 179.39 44.89 10.64 180.64 79.46
52.75 52.21 55.50 50.41 11.22 33.04 11.71 16.15
Gd
Gs
G
4.6 1.9 0.4 0 2.6 5.7 0.0 1.9
4.7 2.9 2.0 0.5 3.0 4.9 2.4 2.9
4.6 2.4 1.2 1 2.8 5 1.2 2.4
* M: mean; SD: standard deviation; Max: maximum; Min: minimum
CONCLUSION Pill identification is a crucial step in evaluating fabric pilling appearance by image analysis. The image of a pilled fabric usually contains image defects, such as uneven background and low contrast, which need to be minimized prior to any feature measurement.
19
Journal of Textile Institute, 88, 488-500, 1997. Floating-yarn points in the image are another barrier for identifying pills, because those points and pills often have similar sizes and intensities. The FFT techniques provide an effective way to separate pills from float-yarn points. Peaks in the power spectrum of the image are the frequency terms resulting from the periodic structure such as floating-yarn points. When the fabric image is reconstructed from the power spectrum, the periodic structure can be removed from the image by deleting peaks in the power spectrum. Pill regions in the non-periodic image can be located by using the template-matching technique and extracted by thresholding the image. Density, size and contrast are the important properties of pills that describe the degree of pilling, and are used as in+dependent variables in the grading equations of pilling.
ACKNOWLEDGEMENT This material is based upon work supported by the National Science Foundation of the United States under Grant No. DMI-9522943, and by the United States Agriculture Department under Grant No. 95-37500-1950.
REFERENCE 1. ASTM D 3511-76, Standard Test method for Pilling Resistance and other Related Surface changes of Textile Fabrics: Brush Pilling Tester Method. 2. ASTM D 3512-76, Standard Test method for Pilling Resistance and other Related Surface changes of Textile Fabrics: Random Tumble Pilling Tester Method. 3. Cooke, W.D., Pilling Attrition and Fatigue, Textile Research Journal, 55, 409-414, 1985. 4. Dougherty, E.R. and Giardina, C.R., “Mathematical Methods for Artificial Intelligence and Autonomous Systems”, Prentice-Hall, Inc., 1988. 5. Jain, A. K., "Fundamentals of Digital Image Processing", Prentice-Hall,Inc. 1989. 6. Konda, A., Xin, L., Takadera, M., Okoshi, Y. and Toricimi, K., J. Text. Mach. Soc. Japan, 36, 96-99, 1988. 7. Krebs, C.J., “Ecological Methodology”, Harper & Row, NY, 1989. 8. Ramgulam, R.B., Amirbayat, J. and Porat, I., The Objective Assessment of Fabric Pilling, Part I: Methodology, J. of Textile Institute, 84, 221-226, 1993. 20
Journal of Textile Institute, 88, 488-500, 1997. 9. Ramgulam, R.B., Amirbayat, J. and Porat, II., The Objective Assessment of Fabric Pilling, Part II: Experimental Work, J. of Textile Institute, 85, 397-401, 1994. 10. Russ, J.C., “The Image Processing Handbook”, CRC Press, 1993. 11. Xu, B. and Y.L. Ting, Fiber Image Analysis: Part I: Fiber Image Enhancement, J. of Textile Institute, in press. 12. Xu, B. and Y.L. Ting, Fiber Image Analysis: Part II: Fiber Characteristics Measurement, J. of Textile Institute, in press. 13. Xu, B., Assessing Carpet Appearance Retention by Image Analysis, Textile Research Journal, 64, 697-709, 1994. 14. Xu, B., Identifying Fabric Structures with Fast Fourier Transform Techniques, Textile Research Journal, in press.
21