A new method for automatic separation of fabric color

Abstract

A novel method combining the characteristics of structure and region information is proposed for automatic segmentation of the color region for different kinds of fabrics. For improving image quality for computer analysis, the structure-texture decomposition processing has been used to extract the main structure from the fabric image, where the fine structure details of fabric yarn patterns have been removed. By using the CIE-Lab color system, the color structure image is then segmented by a fuzzy region-based segmentation model that can be solved efficiently through a fast numerical scheme. The experimental results show that the main disadvantage and difficulty of using color clustering-based methods and commonly used image segmentation methods for fabric color separation is overcome by the proposed method. The proposed method has high accuracy and the computation time is very reasonable. It can be applied to extract fabric color regions for different fabric structures, such as woven, knitted and embroidery structures.

Keywords

fabric texture segmentation color region separation woven fabric embroidery fabric fabric image analysis

Fabric structure and color are two essential parameters in fabric texture analysis. The accuracy and efficiency of structure and color analysis play an important role in product quality control in modern manufacturing technologies, especially with the advent of high-precision weaving and knitting. In the conventional analysis methods, the structure and color parameters, such as fabric weave pattern and the layout of color yarns, are detected manually, which is very tedious and time-consuming work. With the development of computer technology, many researchers have tried to detect fabric parameters via various image-processing techniques.^1–9

Since fabric possesses highly periodic properties originating from the repetition of basic weave patterns or the layout of yarns, researchers had tried to recognize the structure parameters by frequency analysis methods.^2,4,10 The focus of their studies was mainly on the analysis of grayscale fabric images. In actual practice, the visual recognition for colorful fabrics is as important as grayscale fabrics. Research on image-processing techniques and digital color quality control has progressed well in recent years and the use of computers to simulate human vision in performing color analysis is becoming a research focus in the textiles community.^{6,7,9,11–13}

Recent studies on fabric color analysis had used color clustering algorithms to separate colors in the fabric image.^6,7,9 Xin et al.⁶ adopted the gray projection method and an active contour grid model to detect the yarn location only for woven fabrics, in which the image intensity accumulations along warp and weft projections were classified into peak or valley points. The locations of yarns were then obtained according to the classification result of these points. The geometry information of the fabric sample, such as the small deformation and the thickness variations of the yarns, had been detected with the active contour grid model. Cross-point classification was conducted based on the different combinations of the intensities of the yarn edges. A histogram-based color clustering method¹⁴ was then used to separate the color yarns. However, there might be some errors in their automatic clustering results and the manual rectification method was used to correct the clustering results.

Pan et al.⁷ proposed using the Fuzzy C-Means (FCM) clustering method to detect the layout of color yarns in woven fabrics. The cluster validity analysis was used to determine the number of colors. In the spatial domain, the Hough Transform algorithm was adopted to detect the yarn skewness in the fabric image in order to obtain the yarn locations. Their experimental work showed that the proposed method could extract color and structure information from the fabric image. However, the performance of the clustering method was dependent on the cluster validity analysis result, which might fail to figure out the right number of yarn colors, especially for yarns with similar colors in the fabric.

Kuo et al.⁹ investigated different clustering algorithms for embroidery fabric color separation and suggested that the Gustafson–Kessel algorithm was suitable for embroidery fabrics. For image enhancement processing, the median filter and the bilateral filter were used to smooth the fabric texture. The Gustafson–Kessel clustering algorithm was then employed to separate the color regions of the fabric image in the color a* and b* components of the CIE-Lab color system. The experimental result showed that the proposed method could achieve higher accuracy in fabric color separation, compared with the clustering algorithms commonly used in the color separation of color images, such as K-means,¹⁵ K-medoid,¹⁶ FCM¹⁷ and Self-Organizing Map (SOM) clustering algorithm.^5,18

This study aims to explore an effective fabric color separation method that has the advantage of directly separating fabric color in the spatial domain while keeping the sharpest edge response of the main structure. To address the problem of the noise caused by capturing the fabric images, as well as to reduce the influence of yarn textures on the color separation results, this study used the structure-texture decomposition method to enhance the fabric images for color separation. Different from the color clustering methods, a fuzzy region-based segmentation model was developed to separate fabric color regions. The proposed method had been shown to be effective in separating color and texture for different kinds of fabric structures.

Research framework

The technological design process of figured-pattern fabric consists of two main steps. Firstly, a figured-pattern image is separated into different regions for fabric drafting according to the colors of texture patterns. Secondly, a series of woven, knitted or embroidery structures are assigned to these regions and a production file will be generated for different machinery. This study focused on the first step of the design process for region-based color separation to solve the problem of lack of manpower for fabric drafting and to shorten drafting time.

The framework of the proposed method is shown in Figure 1. There were four steps. Firstly, the RGB color image was loaded and processed by the structure-texture decomposition method through which the image was split into two images: (1) the structure image; and (2) the texture image. Secondly, the structure image was transferred into the CIE-Lab color domain. Thirdly, a fuzzy region-based segmentation model was used to separate the color regions in the structure image and these regions were described by region membership functions. Finally, the segmentation result was obtained using an easily adjustable parameter for binarization of the region membership values.

Figure 1.

Research framework of the fabric color separation method.

Image-processing methods

Structure-texture decomposition

Decomposing a color image into meaningful components is a helpful step in fabric image feature detection.¹⁹ Since figured-pattern fabrics, such as Jacquard fabric and embroidery fabric, may have complex pattern and texture in the image, this study used structure-texture decomposition to describe the pattern and texture components. The general concept is that an image can be regarded as composed of a structure part, corresponding to the main large objects in the image, and a textural part, containing fine-scale details, usually with some periodicity and oscillatory nature.²⁰ In fact, such information in fabric images coexists with geometric microstructures, such as yarn contours, shapes and boundaries, or is embedded in coarser structures formed by yarn or dyestuff color variations.

This study proposed a simple and yet effective method to accomplish fabric structure-texture decomposition based on a variational method. Our experimental results, as demonstrated later on in the Experimental details and discussion section, confirmed that structure and texture in the fabric image are completely decomposable. Since enforcing the total variation regularizer to preserve large-scale edges does not require extensive texture information, this study used a structure-texture decomposition model based on relative total variation,²¹ which is expressed as

\arg min \sum_{i, j} (S (i, j) - f (i, j)) 2 + λ \cdot (\frac{D x (i, j)}{L x (i, j) + ɛ} + \frac{D y (i, j)}{L y (i, j) + ɛ})

(1)

where f is the input image, S is the extracted structure image, i and j index two-dimensional (2D) pixels in the image,

f - S

is the resulting texture image, λ is a controlled parameter and ɛ is a small positive number to avoid division by zero. The first term is the data term

(S (i, j) - f (i, j)) 2

, which is to make the input and result not deviate much. The second term is the regularizer term, relative total variation, which includes a general pixel-wise windowed total variation measure

D x (i, j)

and

D y (i, j)

, and a windowed inherent variation, expressed as follows:

D x (i, j) = \sum_{k, l} g (i, j, k, l) \cdot | \partial x S (k, l) |

(2)

D y (i, j) = \sum_{k, l} g (i, j, k, l) \cdot | \partial y S (k, l) |

(3)

L x (i, j) = \sum_{k, l} g (i, j, k, l) \cdot | \partial x S (k, l) |

(4)

L y (i, j) = \sum_{k, l} g (i, j, k, l) \cdot | \partial y S (k, l) |

(5)

where k and l index 2D pixels in the window and g is a weighting function defined according to spatial affinity, expressed as

g (i, j, k, l) \propto \exp (- \frac{(x i, j - x k, l) 2 + (y i, j - y k, l) 2}{2 δ 2})

(6)

where δ is the controlled parameter to adjust the spatial scale of the window.

This study adopted the numerical solution proposed by Xu et al.²¹ for Equation (1). A detailed mathematical study of the numerical solution is given by Xu et al.²¹ and Levin et al.²² The regularizer term can be decomposed into a non-linear term and a quadratic term:

\arg min \sum_{i, j} (S (i, j) - f (i, j)) 2 + λ \cdot (\sum_{k, l} u x (k, l) w x (k, l) \partial x S (k, l) 2 + \sum_{k, l} u y (k, l) w y (k, l) \partial y S (k, l) 2)

(7)

where

u x (k, l) = (G σ * \frac{1}{| G σ * \partial x S | + ɛ}) k, l

w x (k, l) = \frac{1}{| \partial x S (k, l) | + ɛ s}

u y (k, l) = (G σ * \frac{1}{| G σ * \partial y S | + ɛ}) k, l

w y (k, l) = \frac{1}{| \partial y S (k, l) | + ɛ s}

where

G σ

is a Gaussian filter with standard deviation δ, * is the convolution operator and

ɛ s

is a small positive number introduced for numerical stability. The minimization problem can be considered to solve a linear iteration system as follows:

(1 + λ L t) \cdot v_{s}^{t + 1} = v l

(8)

where 1 is an identity matrix,

L t

is the weight matrix computed based on the structural vector

v_{s}^{t}

and v_s and v_l are the vector representation of S and f, respectively. The forward difference method can be used to approximate discrete gradients by solving a linear system with a five-point spatially inhomogeneous sparse Laplacian matrix. Efficient solvers are available for it^22,23 and the numerical experiments for fabric images will be shown in the Experimental details and discussion section.

Fuzzy region-based structure segmentation

Once the structure image was expressed in the CIE-Lab color system, the main large objects should be recognized by the color separation algorithm. To address the problems of image color variations between different regions and blurred shadows of the region boundaries in the structure image, the color separation of the fabric image should be processed by a fuzzy region-based segmentation model. Let Ω be a bounded two-dimension domain and $I (x) : Ω \to R 3$ be a structure image in CIE-Lab color space. We assume that the domain Ω can be partitioned into N sub-regions, $Ω 1, \dots, Ω N$ , along an edge set ${\partial Ω i} i = 1 : N$ , such that $⋃_{i = 1}^{N} Ω i = Ω$ and $Ω i \cap Ω j = Φ$ for $\forall i \neq j, i, j = 1 : N$ . The segmentation model of fuzzy region classification is given by the energy function as follows:

E (U, a) = \sum_{i = 1}^{N} \int Ω | \nabla u i | d x + λ \sum_{i = 1}^{N} \int Ω r_{i}^{a i} u_{i}^{p} d x

(9)

subject to

(i) \sum_{i = 1}^{N} u i = 1, (ii) 0 \leq u i \leq 1, for i = 1 : N

(10)

where N is the number of regions; u is the fuzzy membership function; the term

\int Ω | \nabla u i | d x

equals the perimeter of region

Ω i

;

a = (a 1, \dots, a N)

r_{i}^{a i}

are error functions in region

Ω i

, which equals

| I - c i | 2

, and

a i = c i

are constants; ∇ is the gradient operator; and λ, p are positive parameters to determine the fuzziness of segmentation.

For the efficiency of minimizing the energy equation, we adopted a fast total variation minimization method to solve it^24,25 and took use of Chambolle’s fast dual projection algorithm.²⁶ To this end, auxiliary variables $V = (v 1, \dots, v N)$ are added to approximate E in Equation (9):

E r (U, V, a) = \sum_{i = 1}^{N} \int Ω | \nabla v i | d x + \frac{λ}{p} \sum_{i = 1}^{N} \int Ω r_{i}^{a i} u_{i}^{p} d x + \frac{1}{2 θ} \sum_{i = 1}^{N} \int Ω (v i - u i) 2 d x

(11)

where θ is a parameter that is small enough so that u_i and v_i are approximately identical. In the above equation, fixing U and V, the objective is to minimize:

E 1 (a) = \sum_{i = 1}^{N} \int Ω r_{i}^{a i} u_{i}^{p} d x

(12)

According to the Chan–Vese error function,²⁷ $r_{i}^{a i} = | I - c i | 2$ . The Chan–Vese model assumes that the image can be approximated by the piecewise constant function. In practice, the model is extended to the local case by introducing a symmetric Gaussian kernel function $ω : Ω \to R$ , which satisfies

\int Ω ω (x) d x = 1, ω (- x) = ω (x)

(13)

The error function at $y \in Ω$ is given by

r_{i}^{a i} (y) = \int Ω (I (x) - s i (y)) 2 ω (y - x) d x

(14)

where

a i = s i

is a region parameter function. Taking the derivative of

E 1 (a)

with respect to s_i, Equation (14) becomes

\int Ω u i (x) (I (x) - s i (y)) ω (y - x) d x = 0

(15)

From the above equation, it can be derived that

s i (y) = \frac{\int Ω u i (x) I (x) ω (y - x) d x}{\int Ω u i (x) ω (y - x) d x}

(16)

In Equation (11), fixing u_i and a_i, v_i can be solved by minimizing

\int Ω | \nabla v i | d x + \frac{1}{2 θ} \int Ω (v i - u i) 2 d x

(17)

The fast dual projection algorithm^26,28 is then used to solve the above equation and the solution is shown as

v i = u i - θ div p_{i}^{*}, i = 1 : N

(18)

where the vector

p_{i}^{*}

can be solved by the fixed point method, initializing

p_{i}^{0} = 0

and iterating:

p_{i}^{n + 1} = \frac{p_{i}^{n} + τ \nabla (div p_{i}^{n} - u i / θ)}{1 + τ | \nabla (div p_{i}^{n} - u i / θ) |}

(19)

where τ is small enough to ensure convergence.

In order to solve the membership function, fixing v_i and c_i, we obtain

min E 2 = \frac{λ}{P} \sum_{i = 1}^{N} \int Ω r_{i}^{a i} u_{i}^{p} d x + \frac{1}{2 θ} \sum_{i = 1}^{N} \int Ω (v i - u i) 2 d x

(20)

subject to

(i) $1 - \sum_{i = 1}^{N} u i (x) = 0$ ;

(ii) $1 \geq u i (x) \geq 0, for i = 1 : N$ .

To solve this optimization problem, we set $p = 1$ and relax constraint (i) by the Lagrange multiplier method to minimize

E 2 a = \frac{ε}{2} \int Ω (\sum_{i = 1}^{N} u i - 1) 2 d x + λ \sum_{i = 1}^{N} \int Ω r_{i}^{a i} u i d x + \frac{1}{2 θ} \sum_{i = 1}^{N} \int Ω (v i - u i) 2 d x

(21)

subject to

1 \geq u i (x) \geq 0, for i = 1 : N

where υ is a positive Lagrange multiplier. In order to give an approximate numerical solution, we obtain the optimal condition as follows:

λ r_{i}^{a i} + \frac{1}{θ} (\tilde{u} i (x) - v i (x)) + ε (\sum_{i = 1}^{N} \tilde{u} i - 1) = 0

(22)

The solution of the above equation is

\tilde{u} i = v i - λ θ r_{i}^{a i} - \frac{ε θ (\sum_{j = 1}^{N} (v i - λ θ r_{i}^{a i}) - 1)}{1 + N ε θ}

(23)

By projecting u_i onto $[0, 1]$ , we have

\overset{\land}{u} i : = min {max {u i, 0}, 1}

(24)

where

\overset{\land}{u} i

shows an approximate numerical solution to Equation (21). More details of the segmentation modeling and its various numerical solutions can be found in Han et al.^29,30 and Li et al.²⁵ It will be demonstrated in the Experimental details and discussion section that such an approximate numerical solution with the CIE-Lab color model data can achieve accurate segmentation results for fabric color separation.

Experimental details and discussion

This study discussed the performance of the image-processing steps in the proposed method and made comparison with other methods, such as clustering algorithms and fuzzy region-based segmentation methods. We used MATLAB R2009a language for programming and for integrated design of operational image display. Different image acquisition methods were used in the experimental work. Specifically, three groups of fabric images were investigated in this study, including different collection ways of fabric images. The first group was collected from the Internet. When searching in Google Images, such pictures of fabrics can be found quickly. The second group was acquired from a charge-coupled device (CCD) camera (Canon EOS 550D) and the third group was obtained from a flat scanner (BENQ Scanner, model number 5560c, scan resolution of the vertical and horizontal directions is the same at 600 DPI).

In order to capture the weave textures in the woven fabrics, the high-resolution images of the silk fabrics were obtained by using a special imaging acquisition system. As shown in Figure 2, a CIE Standard Illuminant D65 lamp was mounted above the fabric sample. The illumination system had been chosen in order to perform a repeatable and controlled acquisition able to preserve the colors in the fabrics.³¹ The camera used in this study is a Canon EOS 550D and the lens model is EF-S 18-200 mm. During the experiments, the focal length was set as 100 mm. Other parameters were set as follows: F-stop f/5.6, Exposure time 1/5 s and ISO speed 100. The image resolution to the physical fabric size in our experiments is around 1376 pixels per inch. That is, one inch of the real fabric corresponds to 1376 pixels in the fabric image. The horizontal resolution and the vertical resolution adopted are the same.

Figure 2.

The imaging acquisition system.

After the process of structure-texture decomposition, the fabric structure image was stored in a RGB color model. When the features of the fabric structure define easy distinction in the segmentation or recognizing system, it can be distinguished by different colors, which can avoid mistakes in the recognition system and promote the recognizing rate.^5,7 Compared with RGB model, the CIE-Lab color system approximates human vision and aspires to perceptual uniformity.¹⁷ In fact, while the dependence of the three coordinates on the traditional RGB metric is non-linear,³² the L*, a*, b* metric better facilitates representing texture variation via Gaussian functions in the segmentation model.^30,33 In the experiments, the CIE-Lab color system was thus used for fabric color separation. The color space conversion was conducted as follows.³¹ Firstly, a conversion from the RGB space to the tristimulus values CIEXYZ, under the illuminant D65, was performed. Secondly, the knowledge of the XYZ values was used to obtain the L*, a*, b* values by the conversion method in Kuo and Kao.⁵ For images with unknown illumination systems, the color space conversion process was done the same as that in Han et al.,³⁰ Rao et al.³⁴ and Chen et al.³⁵

Our region-based segmentation method was designed for general color separation of different fabric images. In other words, segmentation performance of the proposed method should be insensitive to imaging resolution and noise cause by image acquisition. In practice, the input fashion of data acquisition can be diverse, including scanning and photography. For this reason, another objective of our method was the robustness of the color analysis method. Therefore, the analysis result should be invariant to the image resolution and insensitive to the data acquisition manner as well as the noise, such as the small fur and the rough texture of the yarn surface. On the other hand, it should be noted that the number of the separated colors depends on the texture contents of the fabric image. In this study, the number of separated colors equals the number of regions to be segmented in the fabric image. In practice, the number of colors should be chosen manually from the corresponding segmentation results. For instance, as demonstrated later on in the experimental results, a fabric image could be segmented into a different number of color regions according to the mixed texture content in some regions of the fabric image.

In the experimental work the fabric image shown in Figure 3 was collected from an embroidery website (http://www.advanced-embroidery-designs.com/). The fabric is classified as an advanced embroidery fabric, where different textures and color variations coexist. We used the color bilateral filter to smoothen the fabric texture in order to make the subsequent color separation more accurate.^9,36 The filtering result is shown in Figure 4, where the image intensities of the background texture and fine details of the fabric have been weakened, while the main structure and the boundaries of different color regions have been kept to a great extent. Next, the structure-texture decomposition method was applied to separate the structure and texture components in the fabric image. As shown in Figure 5, the structure of the fabric image was extracted through the proposed method. It can be seen that the texture details had been completely removed.

Figure 3.

Original fabric image.

Figure 4.

The fabric texture after the processing of color bilateral filtering.

Figure 5.

The structure image obtained from structure-texture decomposition.

After structure-texture decomposition, the fabric image had been split into two images: (1) the structure image; and (2) the texture image. Image intensity maps of these two images should be quite different. In fact, signal components of these images can be examined clearly in a three-dimensional intensity map. Specifically, Figures 6 –8 show the original structure and texture components (mixed components), the extracted structure component and the extracted texture component, respectively. It should be noted that the corresponding images are gray intensities for the convenience of display. In the center of the images, it can be found that the petals of the flower in Figure 7 have a more uniform structure than those in Figure 6. The experimental result in Figure 7 showed that different regions indicate different contents in the structure image and these regions exhibit different intensity properties in the spatial domain, making them more easily decomposable.

Figure 6.

The intensity map of the original fabric image.

Figure 7.

The intensity map of the extracted structure image.

Figure 8.

The intensity map of the extracted texture image.

Once the structure image had been obtained from the structure-texture decomposition method, the image was converted into the CIE-Lab color domain. The color components L*, a* and b* were then used for fuzzy region-based image segmentation. The output of the segmentation model was u in Equation (11). For a specific region i, the membership values could be displayed in a grayscale image. For instance, one region membership function $u i (x, y)$ in the embroidery fabric image is shown in Figure 9 in a grayscale format and its membership values are shown in Figure 10, where the white region indicates the segmented region i. The black region indicates that the region is not segmented into region i and the gray region indicates that the region may be segmented into region i. In Figure 10, the membership values in the region membership function can be seen in an intuitive way; the corresponding binary values are shown in Figure 11. In all experiments, the final segmentation results were given after the binarization processing and the threshold value was 0.5. The result showed that the pixels inside the region had the same membership values with their neighbor pixels and the membership values of the pixels near the boundaries were dependent on the threshold value.

Figure 9.

Illustration of a region membership function. Note that the middle image is the enlarged part of the left image and the right image is the corresponding enlarged part of the structure image.

Figure 10.

The membership values in the region membership function shown in Figure 9.

Figure 11.

The binary values of the membership values in Figure 10.

A three-region segmentation result of the fabric image is shown in Figure 12. It shows that the fabric image can be segmented into several different color regions and each region represents different objects in the spatial domain. From the edge map in Figure 12, it can be seen that the boundaries of the segmented regions are smooth and the classification result of main large objects agrees with user observation. Furthermore, in order to separate flower color in Figure 12(c), a new segmentation process was conducted only for the flower region. The target of this segmentation process was to separate the pistil and petal regions. The final segmentation result is shown in Figure 13, where the regions of the original fabric image are given in the final segmental result. The corresponding binary result is shown in Figure 14, where the regions can be directly used for structured texture assignment during the technical design process.

Figure 12.

The fuzzy region-based segmentation results: (a) is the background fabric texture; (b) is the leaves of the image; (c) is the flower of the image; (d) is the edge map.

Figure 13.

The final segmentation result of original fabric image.

Figure 14.

The binarization result of the segmented regions.

From the experiments shown in Figures 13 and 14, it can be found that the proposed method for structure-texture decomposition was suitable for dealing with different fabric textures. The experiments showed that the reliability and the performance of the structure-texture decomposition method were excellent when compared with those of the traditional texture filtering techniques, for example, the bilateral filters. Specifically, after the decomposition process the fine details of yarn texture, such as the long floats of individual yarns and the interstices of interwoven yarns, had been completely removed and the main structure could be segmented correctly by the fuzzy region segmentation model. However, it should be noted that the main constraint of the method is that some local boundaries were lost after the processing. For instance, some of the light-green boundaries (the binding threads around the leaves in the sewing process) of the leaves disappeared and were incorporated into one region. Therefore, in the industry the decomposition process should be implemented in a local fashion to avoid the loss of the interest of fine structures in some local boundaries. In practice, an interactive way for adjusting the extent of the decomposition could be applied to conduct texture or structure separation under human supervision. The processing results could be used for further analysis of fabric texture or structure.

In order to evaluate the effectiveness of the proposed method, comparison with other methods was made. Five other methods were considered: the FCM clustering method,⁷ the Gustafson–Kessel clustering method,⁹ the Expectation-Maximization segmentation method,³⁷ the traditional fuzzy region-based segmentation method²⁵ and the PCA-based fuzzy region competition method.³⁰ The experimental details are shown in Figure 15. The color separation results of different methods are different and the result of our method is the best. In the experimental results, it can be found that there are many tiny regions of color clustering methods, especially in the FCM clustering method. The reason for this is that the color clustering methods are not able to measure the spatial affinity of the neighboring pixels during the color separation process.

Figure 15.

Fabric color separation results of different methods. From top to bottom, the first row, the second row, the third row, the fourth row, the fifth row and the sixth row show the segmented color regions of the fabric image by the Fuzzy C-Means clustering method, the Gustafson–Kessel clustering method, the Expectation-Maximization segmentation method, the traditional fuzzy region-based segmentation method, the Principal component analysis (PCA)-based fuzzy region competition method and the proposed method, respectively.

On the other hand, our study confirmed that the color separation result of the Gustafson–Kessel clustering method is better than that of the FCM clustering method, because the data set of the former has data set distribution information, in addition to the location of the center of gravity.⁹ The experimental results suggest that the technologies commonly used in color image color separation may not be suitable for fabric color separation, for example the Expectation-Maximization segmentation method³⁷ and the traditional fuzzy region-based segmentation method.⁹ The experimental results also show that there may be some unexpected results of these methods in that the fabric image has different weaving methods to create different textures, which may affect the color separation results. Although the PCA-based fuzzy region competition method³⁰ can achieve better color separation quality than the previous methods, it fails to separate fabric color where the color variation exists in the color regions. For instance, some regions of the leaves have been misclassified into the background texture due to the color variation inside the leaves.

This study also used different fabric structures for evaluating the proposed method. As shown in Figures 16 –21, the experimental results confirmed the effectiveness of the proposed method for fabric color separation. The fabric structures in the experimental work included embroidery fabric, woven fabric and knitted fabric. The results showed that the structure-texture decomposition method was able to remove the texture from the fabric image. In the experiments, the fabric images have multiple color regions. For instance, Figure 16 shows an embroidery fabric image with five color regions and our method could separate the colors correctly. During the binarization processing of the fuzzy region membership function, the threshold value should be very easy to choose when the number of separated colors is chosen correctly. The range of the threshold value is the range of the membership value $[0, 1]$ and in all experiments the threshold value was fixed as the medium value (0.5). The experiments showed that the main structure of all fabric images could be segmented correctly by using the medium value. The reason to choose the medium value is that without prior knowledge the binarization processing of the membership values of the neighboring regions should be neutral. However, the hairiness of interwoven yarns may exist in some local regions of the fabric texture and it is difficult to separate some pixels near the boundaries of these regions. Therefore, the best threshold value for these local regions should be determined manually according to the segmentation results.

Figure 16.

Five-region color separation results for an embroidery fabric image with flower figured-pattern. (a) Original fabric image. (b) The structure image. (c)–(g) The segmented color regions. (h) The edge map. The embroidery fabric image was found by searching the website of Advanced Embroidery Designs (http://www.advanced-embroidery-designs.com/).

Figure 17.

Three-region color separation results for a woven fabric image with geometrical figured-pattern (left) and an embroidery fabric image with fish figured-pattern (right). Left: (a) Original fabric image. (b) The structure image. (c)–(e) The segmented color regions. (f) The edge map. Right: (a) Original fabric image. (b) The structure image. (c)–(e) The segmented color regions. (f) The edge map. The first fabric image was found by searching the website of Houzz and the second fabric image was found by searching the website of Advanced Embroidery Designs.

Figure 18.

Color separation results for a woven fabric image with flower figured-pattern (left) and an embroidery fabric image with flower figured-pattern (right). Left: (a) Original fabric image. (b) The structure image. (c)–(e) The segmented color regions. (f) The edge map. Right: (a) Original fabric image. (b) The structure image. (c)–(g) The segmented color regions. (h) The edge map. The left fabric image was found by searching the website of Houzz and the right fabric image was found by searching the website of Advanced Embroidery Designs.

Figure 19.

Two-region color separation results for an embroidery fabric image with geometrical figured-pattern (left) and a knitting fabric image with geometrical figured-pattern (right). Left: (a) Original fabric image. (b) The structure image. (c)–(d) The segmented color regions. (e) The edge map. Right: (a) Original fabric image. (b) The structure image. (c)–(d) The segmented color regions. (e) The edge map. The left fabric image was found by searching the website of Houzz and the right fabric image was found by searching the website of Advanced Embroidery Designs.

Figure 20.

Color separation results for woven fabric images that were acquired by a camera (Canon EOS 550D). Left: (a) Original fabric image. (b) The structure image. (c)–(d) The segmented color regions. (e) The edge map. Right: (a) Original fabric image. (b) The structure image. (c)–(e) The segmented color regions. (f) The edge map. The woven fabrics were collected from a Jacquard company.

Figure 21.

Color separation results for woven fabric images that were acquired by a flat scanner (BENQ Scanner 5560c). Left: (a) Original fabric image. (b) The structure image. (c)–(d) The segmented color regions. (e) The edge map. Right: (a) Original fabric image. (b) The structure image. (c)–(e) The segmented color regions. (f) The edge map. The woven fabrics were collected from a Jacquard company.

For these fabric images and the corresponding segmentation results shown in Figures 16 –21, only structure extraction and region segmentation had been applied to fabric color separation in the CIE-Lab color domain. It should be noted that the commonly used preprocessing techniques for noise reduction or texture smoothing, such as the bilateral filter and the median filter, had not been used in these experiments. Nevertheless, it can be seen that the proposed method was still able to separate fabric color for these fabric images and the results were consistent with the expectation of human eyes. In our experimental work, the parameters in the segmentation process were fixed: $λ = 5$ and $θ = 0.01$ . The computation time of the structure-texture decomposition process and the segmentation process for a fabric sample was about 3 and 20 s, respectively. All the fabric images were resized and the number of row and column pixels was specified as $256 \times 256$ . After the segmentation process, the segmented results were labeled by the binary values and different color regions were expressed by the binary value matrices, which were resized into the original size in order to map the original image. The image resizing process was carried out by the Bicubic interpolation method³⁸ in which the output pixel value is a weighted average of pixels in the nearest 4-by-4 neighborhood. Note that the central processing unit (CPU) of the machine is an Intel Core2 Duo T9400, 2.53 GHz. It showed that the proposed method for fabric color separation was of low computation cost.

Conclusions

In this paper a fuzzy region-based segmentation method for automatic fabric color separation was described. The proposed image-processing approach offered an effective analysis for separating fabric color according to the color values in different regions. The materials used in our experimental work included a wide variety of fabrics with different target regions. Different image acquisition methods were used to obtain fabric images with different structures, including woven structure, embroidery structure and knitted structure. The types of experimental samples exceeded those in previous studies. The experiments showed that the proposed method was accurately able to segment fabric color regions along their boundaries. Moreover, the tiny regions had been incorporated into the large regions so that the main large objects could be recognized and the main structure had been preserved. The preliminary experiments confirmed the effectiveness of our method.

Successful separation of fabric color by the proposed approach was mainly the result of the following three aspects. Firstly, the fabric image had been split into two components, structure and texture, and the influence of the texture component on the segmentation process had been eliminated. Secondly, in the structure domain, the image was expressed in the CIE-Lab color system, which describes the color components better than RGB color system. Thirdly, a multi-region segmentation model was used to extract the color regions in the structure image and the fuzzy membership function was used to express the blurred shadows and boundaries of color regions.

This paper showed that the proposed method was effective in separating fabric color and the computation cost was low. The experiments showed that, by using the structure-texture decomposition processing, fabric images that had not been processed by the commonly used filtering techniques, such as the bilateral filter and the median filter, still could been segmented accurately by the fuzzy region-based segmentation model. The potential applications of our method could be found in fabric texture design or content-based automatic evaluation of fabric images.

Footnotes

Funding

This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

References

Wood

. Applying Fourier and associated transforms to pattern characterization in textiles. Text Res J 1990; 60: 212–220.

. Identifying fabric structures with fast Fourier transform techniques. Text Res J 1996; 66: 496–506.

Millan

Escofet

. Fourier domain-based angular correlation for quasiperiodic pattern recognition. J Appl Opt 1996; 35: 6253–6260.

Lin

. Applying a co-occurrence matrix to automatic inspection of weaving density for woven fabrics. Text Res J 2002; 72: 486–490.

Kuo

Kao

. Self-organization map network for automatically recognizing color texture fabric nature. Fibers Polym 2007; 8: 174–180.

Xin

Baciu

. Investigation on classification of weave pattern based on an active grid model. Text Res J 2009; 79: 1123–1134.

Pan

Gao

Liu

. Automatic detection of the layout of color yarns for yarn-dyed fabric via a FCM algorithm. Text Res J 2010; 80: 1222–1231.

Zheng

Han

. A new method for classification of woven structure for yarn-dyed fabric. Text Res J 2014; 84: 78–95.

Kuo

Jian

. Automatic machine embroidery image color analysis system. Part I: Using Gustafson-Kessel clustering algorithm in embroidery fabric color separation. Text Res J 2012; 82: 571–583.

10.

Kuo

Tsai

. Automatic recognition of fabric nature by using the approach of texture analysis. Text Res J 2006; 76: 375–381.

11.

Pan

Liu

Gao

. Measuring linear density of threads in single-system-mélange color fabrics with FCM algorithm. Color Res Appl 2012; 38: 456–462.

12.

Zheng D, Han Y, Baciu G, et al. Design through cognitive color theme: a new approach for fabric design. In: 11th IEEE international conference on cognitive informatics & cognitive computing, 2012, pp.346-355.

13.

Han

Zheng

Baciu

. Fuzzy region competition-based auto-color-theme design for textile images. Text Res J 2013; 83: 638–650.

14.

Shafarenko

Petrou

Kittler

. Histogram-based segmentation in a perceptually uniform color space. IEEE Trans Image Process 1998; 7: 1354–1358.

15.

MacQueen

Some methods for classification and analysis of multivariate observations. In: LeCam

Neyman

(eds). proceedings of the fifth Berkeley symposium on mathematical statistics and probability, Berkeley, CA: University of California Press, 1967, pp. 281–297.

16.

Kaufman

Rousseeuw

. Finding groups in data: An introduction to cluster analysis, New York: Wiley, 1990.

17.

Kuo

Shih

Lee

. Automatic recognition of fabric weave patterns by a fuzzy c-means clustering method. Text Res J 2004; 74: 107–111.

18.

Furferi

Governi

. The recycling of wool clothes: An artificial neural network colour classification tool. Int J Adv Manuf Technol 2008; 37: 722–731.

19.

Zheng

. A new method for the detection and classification of weave pattern repeat. Text Res J 2014; 84: 1586–1599.

20.

Aujol

Gilboa

Chan

. Structure-texture image decomposition—modeling, algorithms, and parameter selection. Int J Comput Vision 2006; 67: 111–136.

21.

Yan

Xia

. Structure extraction from texture via relative total variation. ACM Trans Graph 2012; 31: 1310–1391.

22.

Levin

Lischinski

Weiss

. Colorization using optimization. ACM Trans Graph 2004; 23: 689–694.

23.

Krishnan

Szeliski

. Multigrid and multilevel preconditioners for computational photography. ACM Trans Graph 2011; 30: 177–177.

24.

Huang

Wen

. A fast total variation minimization method for image restoration. SIAM J Multiscale Model Simulat 2008; 7: 774–795.

25.

Zeng

. A multiphase image segmentation method based on fuzzy region competition. SIAM J Imag Sci 2010; 3: 277–299.

26.

Chambolle

Lions

. Image recovery via total variation minimization and related problems. Numer Math 1997; 76: 167–188.

27.

Chan

Vese

. Active contour without edges. IEEE Trans Image Process 2001; 10: 266–277.

28.

Chambolle

. An algorithm for total variation minimization and applications. J Math Imaging Vision 2004; 20: 89–97.

29.

Han

Wang

Feng

. A fast multiphase image segmentation algorithm based on nonconvex regularizer. Pattern Recognit 2012; 45: 363–372.

30.

Han

Feng

Baciu

. Variational and PCA based natural image segmentation. Pattern Recognit 2013; 46: 1971–1984.

31.

Furferi

. Colour classification method for recycled melange fabrics. J Appl Sci 2011; 11: 236–246.

32.

Comaniciu

Meer

. Mean shift: A robust approach toward feature space analysis. IEEE Trans Pattern Anal Mach Intell 2002; 24: 603–619.

33.

Yang A, Wright J, Sastry S, et al. Unsupervised segmentation of natural images via lossy data compression. EECS Department, University of California, Berkeley, 2006, pp.1--22.

34.

Rao S, Mobahi H, Yang A, et al. Natural image segmentation with adaptive texture and boundary encoding (Lecture Notes in Computer Science, vol. 5994). New York: Springer, 2009; pp. 135--146.

35.

Chen

Cao

Wang

. Image segmentation by MAP-ML estimations. IEEE Trans Image Process 2010; 19: 2254–2264.

36.

Tomasi C and Manduchi R. Bilateral filtering for gray and color images. In: IEEE sixth international conference on computer vision, 1998, pp. 839-846.

37.

Carson

Belongie

Greenspan

. Blobworld: Image segmentation using expectation-maximization and its application to image querying. IEEE Trans Pattern Anal Mach Intell 2002; 24: 1026–1038.

38.

Keys

. Cubic convolution interpolation for digital image processing. IEEE Trans Signal Proc Acoustics Speech Signal Proc 1981; 29: 1153–1160.