Colored spun fabric texture representation and application by combining spatial features with frequency features

Abstract

Colored spun fabrics are difficult to accurately characterize with a local binary pattern due to texture anisotropy caused by the uneven distribution of dyed fibers. In this paper, we present a texture representation model based on spatial and frequency characteristics. The proposed model takes advantage of the local binary pattern and local phase quantization to extract the texture of woven fabric. Then, the two features are connected in series, and the features of dimension reduction by principal component analysis are used to represent the texture of the fabric image. Finally, the hierarchical hybrid classifier is applied to classify the fabric structure. The experimental results show that the local phase quantization feature is robust to the fuzzy transformation and the texture representation model has a stronger ability of texture description than the single local binary pattern feature, with the average classification accuracy of 97.59% on 336 samples. In addition, compared with the deep learning algorithm, the texture representation algorithm can ensure a high classification accuracy.

Keywords

colored spun fabric texture representation local binary pattern local phase quantization hierarchical hybrid classifier fabric structure classification

Woven fabric is composed of warp and weft yarns interweaved vertically according to certain rules, and its texture has certain periodicity, directionality and structure characteristics.¹ Effective extraction or representation can not only promote fabric quality index evaluation, but also lay a foundation for fabric quality detection. At present, the application of fabric texture representation mainly involves two aspects:^2–4 (a) objective measurement of the geometric or statistical characteristics of the fabric surface texture, such as the grade estimation of the fabric wrinkles, the detection of the fabric density, the automatic identification of the fabric weave, etc.; (b) abnormal analysis of fabric surface texture features, such as automatic detection of fabric defects.

Common texture representation methods can be divided into statistical methods, structure methods, model methods, filtering methods and deep learning methods.^5,6 Combining with different structural features and application requirements of textiles, many research institutions and teams at home and abroad have conducted extensive research on texture representation algorithms of fabric images.

Due to the advantages of the linear binary pattern (LBP) in texture representation, a large number of LBP optimization and improvement algorithms have emerged in recent years. These methods can be roughly classified into three categories:^7–9 (a) changing coding or mode selection; (b) changing the domain topology or sampling structure; (c) combining the LBP with other complementary features. Pawening et al.¹⁰ proposed a textile image classification algorithm based on the linear binary pattern and gray-level co-occurrence matrix (GLCM). Compared with the single feature, the fusion feature information had the best accuracy. Tajeripour et al.¹¹ made use of modified LBPs to conduct defect detection in patterned fabrics. This method was effective for testing star patterned fabrics, dot patterned fabrics and box patterned fabrics with a detection rate of over 95%. Liu et al.¹² applied the main local binary pattern (MLBP) to reconstruct the fabric texture, and achieved defect detection. Li et al.¹³ converted the fabric image from RGB color space to LAB color space and extracted the LBP operator of the energy-based feature images to defect defects with a detection success rate of more than 94.09%. Zhang et al.¹⁴ put forward local statistics and global analysis to achieve the effective representation of fabric texture.

Unlike a monochromatic object, colored spun fabrics use dyed fibers as the basic carrier of color, as shown in Figure 1. In the process of yarn forming or weaving, the dyed fiber appears as a spiral related to the twist on the surface of yarn or fabric and the fibers stack and gather with each other. Hence, this heterogeneity and anisotropy destroy the original periodicity, directivity and structural characteristics of the colored spun fabric, resulting in fabric texture image blur and obvious noise.¹⁵ However, the LBP algorithm can only extract the spatial features of the image, that is, process the gray-level changes between image pixels, and cannot reflect the gradient changes between pixels, that is, the frequency information, which has certain limitations.

Figure 1.

Colored spun fabrics.

In order to solve this problem, a texture representation model of colored spun woven fabric is proposed in this paper, which combines spatial and frequency features. Firstly, the feature values are extracted by the LBP and local phase quantization (LPQ) from woven fabric images, respectively. Then, the serial feature of the LBP and LPQ is optimized by applying principal component analysis (PCA) to reduce the dimension. Finally, the optimized texture features are used as the input of a hierarchical hybrid classifier to classify colored spun woven fabrics.

The structure of this paper is organized as follows. In the second section, we introduce the research method in detail. In the third section, we give experiments and results, including experiment preparation, parameter optimization, experimental results and comparison experiments. Conclusions follow in the last section.

Research method

LBP algorithm

The LBP is a classical algorithm used to describe the local texture features of images. Its basic principle is to compare the gray values of the central pixel and the domain pixel. If the gray value of a domain pixel is greater than or equal to the gray value of the central pixel, it is set to 1; otherwise, it is set to 0. The rules are shown in the following equations

{LBP}_{P, R} = \sum_{i = 0}^{P - 1} s (g_{i} - g_{c}) \times 2^{i}

(1)

s (x) = {\begin{matrix} 1 & x \geq 0 \\ 0 & else \end{matrix}

(2)

In Equation (1), g_c is the gray value of the center pixel, g_i is the gray value of the pixels in its domain, P represents the number of surrounding pixels and R represents the domain radius of the center. In order to further improve the applicable range of the LBP operator, it can be extended to the LBP algorithm with a variable region, and the circle field can be employed to replace the square field.¹⁶

LPQ algorithm

The LPQ algorithm uses short-time Fourier transform (STFT) to calculate the local phase information of pixels on the image.¹⁷ The image $f (x, y)$ adopts the discrete STFT, and the calculation method is as follows

F (u, x) = \sum_{y \in Nx} f (x - y) e^{- j 2 π u^{T} y}

(3)

where

Nx

represents a neighborhood with a size of

M \times M

, the local Fourier coefficients are calculated by four frequency points

u_{1} = [a 0] T

u_{2} = [0 a] T

u_{3} = [a a] T

and

u_{4} = [a - a] T

and

a = \frac{1}{M}

represents a small range. Hence, pixels are represented by vector

F_{x}^{c} [F (u_{1}, x), F (u_{2}, x), F (u_{3}, x), F (u_{4}, x)]

. The phase of the Fourier coefficient can be represented by the symbols of the real number and the imaginary number of each part, as shown in the following equation

q_{j} = {\begin{matrix} 1 & g_{j} \geq 0 \\ 0 & g_{j} < 0 \end{matrix}

(4)

where g_j is the jth part of the vector

G (x) = [Re {F (x)} Im {F (x)}]

. Then, q_j is encoded in binary as follows

f_{LPQ} (x) = \sum_{j = 1}^{8} q_{j} \times 2^{j - 1}

(5)

where

LPQ (x)

represents the algorithm with window size of

M \times M

, and the specific parameters need to be optimized according to the texture features of the sample image.

Fusion of LBP and LPQ features

No matter whether in the spatial domain or the frequency domain, some information is always lost in the texture analysis. The combination of the two can complement each other and enhance the ability of texture description. Therefore, the texture image and histogram features of the colored spun woven fabric extracted by the LBP and LPQ are shown in Figure 2.

Figure 2.

Colored spun woven fabric texture features in spatial and frequency domains: (a) original picture of colored spun woven fabric; (b) local binary pattern (LBP) pseudo grayscale of colored spun woven fabric; (c) local phase quantization (LPQ) pseudo grayscale of colored spun woven fabric; (d) LBP histogram features; (e) LPQ histogram features.

It can be seen from Figure 2 that there are significant differences between the LBP and LPQ texture features of the plain weave, that is, the ability to depict texture is different and complementary. The fusion of the two features can improve the ability of texture feature descriptors to represent complex modes. The details are as follows

Feature = PCA ({Hist}_{LBP} + {Hist}_{LPQ})

(6)

PCA is a multivariate statistical method, which transforms a set of related variables into a set of orthogonal and unrelated variables through orthogonal transformation. The transformed variables are called principal components. Its goal is to extract the most important information from the data set and compress the size of the data set by reducing the dimensions without losing too much information. Here, $PCA (•)$ can be used to reduce the high feature dimension. $Hist (•)$ represents the histogram feature, and ${Hist}_{LBP} + {Hist}_{LPQ}$ represent the serial features of the LBP and LPQ, respectively.

Experiments and results

Experimental preparation

There are many factors that can affect the texture change of colored spun woven fabrics, including fibrous materials, quality ratio of dyed fibers, color of dyed fibers, warp and weft density and fabric structure. According to the requirements of the experiment, a company was commissioned to prepare 56 samples of colored spun woven fabrics. Table 1 shows the dyeing fiber blending parameters, twist factor of the colored spun yarn and fabric structure. The quality ratio of dyed fibers varied from 0.5% to 4.0% in the first batch of experimental samples.

Table 1.

Parameters of the first batch of experimental samples for colored spun woven fabric

Sample number	Twist factor	Quality ratio of dyed fibers (%)
SS41A	330	White 30.00: red 35.00: green 35.00
SS43A	370	White 30.00: red 35.00: green 35.00
SS51A	330	White 30.00: red 34.75: green 35.25
SS53A	370	White 30.00: red 34.75: green 35.25
SS61A	330	White 30.00: red 34.50: green 35.50
SS63A	370	White 30.00: red 34.50: green 35.50
SS71A	330	White 30.00: red 34.00: green 36.00
SS73A	370	White 30.00: red 34.00: green 36.00
SS81A	330	White 30.00: red 33.00: green 37.00
SS83A	370	White 30.00: red 33.00: green 37.00
SS41D	330	White 30.00: red 35.00: green 35.00
SS43D	370	White 30.00: red 35.00: green 35.00
SS51D	330	White 30.00: red 34.75: green 35.25
SS53D	370	White 30.00: red 34.75: green 35.25
SS61D	330	White 30.00: red 34.50: green 35.50
SS63D	370	White 30.00: red 34.50: green 35.50
SS71D	330	White 30.00: red 34.00: green 36.00
SS73D	370	White 30.00: red 34.00: green 36.00
SS81D	330	White 30.00: red 33.00: green 37.00
SS83D	370	White 30.00: red 33.00: green 37.00
SS41F	330	White 30.00: red 35.00: green 35.00
SS43F	370	White 30.00: red 35.00: green 35.00
SS51F	330	White 30.00: red 34.75: green 35.25
SS53F	370	White 30.00: red 34.75: green 35.25
SS61F	330	White 30.00: red 34.50: green 35.50
SS63F	370	White 30.00: red 34.50: green 35.50
SS71F	330	White 30.00: red 34.00: green 36.00
SS73F	370	White 30.00: red 34.00: green 36.00
SS81F	330	White 30.00: red 33.00: green 37.00
SS83F	370	White 30.00: red 33.00: green 37.00

In Table 1, the first batch of samples is woven from white, red and green polyester fibers, and we adopt the spinning method of ring spinning. The length of the dyed fiber is 38 mm, the linear density is 1.65 tex, the twist direction of double-ply yarn is S and the quality ratio ranges from 0.5% to 4.0%. At the same time, the fabric weave adopts the plain, twill and satin weave, which are called the three elementary weaves. The specific weaving process parameters and number sequences are shown in Table 2.

Table 2.

Weaving process parameters of the first batch of colored spun woven fabric samples

Pattern weave	Number sequence	On-loom weft density (root/10 cm)	Off-loom warp density (root/10 cm)	Off-loom weft density (root/10 cm)	Reed number
Plain	SSxxA	230	380	226	92
Twill	SSxxD	300	460	300	110
Stain	SSxxF	330	488	334	92

The second batch of 26 samples is woven from white, red and yellow cotton fibers and white, blue and black cotton fibers, respectively. The samples are plain weave with 130 reeds, and the off-loom weft density is 280/(10 cm). The ratio of dyed fibers among samples varies randomly in a wide range; the specific parameters are shown in Table 3.

Table 3.

Parameters of the second batch of colored spun woven fabric

Sample number	Twist factor	Quality ratio of dyed fibers (%)
17001	350	White 95.05: red 3.93: yellow 1.02
17002	350	White 94.60: red 3.90: yellow 1.48
17003	70	White 94.10: red 3.90: yellow 2.00
17004	350	White 93.60: red 3.90: yellow 2.50
17005	350	White 93.10: red 3.90: yellow 3.00
17006	350	White 94.96: red 3.06: yellow 1.98
17007	350	White 94.51: red 3.52: yellow 1.97
17008	350	White 94.08: red 3.90: yellow 2.02
17009	350	White 93.38: red 4.60: yellow 2.02
17010	350	White 92.90: red 5.10: yellow 2.00
17011	350	White 93.90: red 4.60: yellow 1.50
17012	350	White 94.00: red 3.53: yellow 2.47
17014	350	White 93.75: red 4.70: yellow 1.55
17015	350	White 94.01: red 3.53: yellow 2.46
17018	350	White 92.00: black 4.00: blue 4.00
17019	350	White 90.00: black 4.00: blue 6.00
17020	350	White 88.00: black 4.00: blue 8.00
17021	350	White 90.00: black 2.00: blue 8.00
17022	350	White 89.00: black 3.00: blue 8.00
17023	350	White 88.10: black 4.00: blue 7.90
17024	350	White 91.00: black 3.00: blue 6.00
17025	350	White 92.00: black 2.00: blue 6.00
17029	60	White 94.30: red 3.76: yellow 1.94
17030	65	White 94.00: red 4.00: yellow 2.00
17031	75	White 94.00: red 4.00: yellow 2.00
17032	80	White 94.30: red 3.80: yellow 1.90

After all samples are balanced at a relative humidity of 65%, we make use of the DigiEye Digital Imaging System to acquire images, and perform white balance and color correction on the DigiEye system camera through whiteboard and standard color card before acquisition. Then, eight standard images of different areas are acquired for each fabric sample and are segmented to obtain 896 sample images of 256 × 256 pixels. Finally, 560 images are chosen as the training samples and the rest as testing samples. Some samples are shown in Figure 3.

Figure 3.

Some experimental samples of colored spun woven fabric: (a) SS41A; (b) SS51D; (c) SS41F; (d) 17001; (e) 17018.

The experimental operation platform is the ultra micro 7048GR-TR deep learning system, Windows 10 operating system, Intel(R) Xeon(R) E5-2678V3*2 CPU, 128 G RAM and NVIDIA Tesla M40*2 operation card. The algorithm development environment is MATLAB 2015b and Python3.7, and TensorFlow-gpu1.14.0, CUDA Toolkit10.0 and cuDNN7.4 are installed.

Parameter optimization

Automatic classification of fabric weave is a typical application of texture representation. In this paper, referring to the construction process of the hierarchical hybrid classifier in Xue and Titterington,¹⁸ the texture features of the first batch of colored spun woven fabric are classified so as to quantitatively analyze the effectiveness and stability of the texture representation model. At the same time, the parameters of the representation model are optimized by combining them with the texture features of the analysis object. This mainly includes the variable area size R of the LBP operator, the number p of pixels in the area, the window size M of the LPQ operator and the size N of the image block. The specific process is as shown in Table 4.

Table 4.

Fabric weave classification results under different parameters of the local binary pattern operator

Pattern weave	Classification accuracy of fabric structure (%)
Pattern weave	${LBP}_{P, R}$ ( $P = 8, R = 1$ )	${LBP}_{P, R}$ ( $P = 8, R = 2$ )	${LBP}_{P, R}$ ( $P = 16, R = 2$ )
Plain	75.61	81.36	80.71
Twill	72.23	78.85	79.45
Stain	76.37	80.21	80.79

The results show that the parameter

P, R

of the uniform rotation-invariant LBP, that is, the number of domain pixels and the variable region size of the operator, will affect the classification result. At

P = 8, R = 2

, the texture representation ability and operation time of the operator can reach the best performance. However, generally speaking, the effect of texture feature extraction and fabric structure classification for the original sample image is not ideal. As a result, the LBP features of the block statistical image can be used for classification; the results are shown in Table 5.

Table 5.

Classification results of image block size and fabric structure

Pattern weave	Classification accuracy of fabric structure (%) ${LBP}_{8, 2}$
Pattern weave	Image block ( $N = 128$ )	Image block ( $N = 64$ )	Image block ( $N = 32$ )
Plain	88.28	88.86	74.28
Twill	86.52	84.52	79.52
Stain	85.86	87.86	80.86

The results show that the classification accuracy can be improved effectively by segmenting the original sample image. At

N = 128

and

N = 64

, the average classification accuracy reaches 87.88% and 87.08%, respectively. However, at

N = 32

, the average classification accuracy is only 78.22% due to the excessive image blocks, which leads to the destruction of spatial information of the fabric structure. At the same time, because of image segmentation, the final texture feature dimension will increase significantly. In addition, the window size of the LPQ operator also affects the expression of gradient information in the frequency domain. Therefore, the classification results of the LPQ algorithm with different window sizes are as shown in Table 6.

Table 6.

Window size of the local phase quantization operator and fabric structure classification results

Pattern weave	Classification accuracy of fabric structure(%) ${LPQ}_{M}$
Pattern weave	Window size ( $M = 5$ )	Window size ( $M = 7$ )	Window size ( $M = 9$ )
Plain	82.35	86.97	85.98
Twill	80.65	83.68	81.45
Stain	81.92	84.68	84.21

The results show that LPQ has the best ability of texture representation and the highest classification accuracy at $M = 7$ . Therefore, based on classification accuracy and computational efficiency, this paper uses $N = 128$ to block the sample image, and then uses the uniform rotation-invariant ${LBP}_{8, 2}$ and ${LPQ}_{7}$ operators to jointly characterize the fabric texture features.

Experimental results and analysis

In order to verify the effectiveness and robustness of the method, the paper analyzes different influencing factors through experiments. When the twist coefficient of the colored spun yarn remains the same, the classification results of the colored spun fabric structure with different quality ratios of the dyed fibers are as shown in Tables 7 and 8.

Table 7.

Classification results with twist factor of 330

Pattern weave	Sample number	Classification accuracy (%)	Average classification accuracy (%)
Plain	SS41A	100.00	100.00
	SS51A	100.00
	SS61A	100.00
	SS71A	100.00
	SS81A	100.00
Twill	SS41D	100.00	100.00
	SS51D	100.00
	SS61D	100.00
	SS71D	100.00
	SS81D	100.00
Stain	SS41F	93.33	95.99
	SS51F	100.00
	SS61F	100.00
	SS71F	93.33
	SS81F	93.33

Table 8.

Classification results with twist factor of 370

Pattern weave	Sample number	Classification accuracy (%)	Average classification accuracy (%)
Plain	SS43A	100.00	100.00
	SS53A	100.00
	SS63A	100.00
	SS73A	100.00
	SS83A	100.00
Twill	SS43D	100.00	94.73
	SS53D	100.00
	SS63D	73.65
	SS73D	100.00
	SS83D	100.00
Stain	SS43F	100.00	93.11
	SS53F	65.55
	SS63F	100.00
	SS73F	100.00
	SS83F	100.00

The experimental results show that the changes in the blending ratio parameters and twist coefficients of the dyed fibers will affect the coloration of the colored spun fabrics. When the twist coefficient is 330, the classification accuracy of plain and twill weave is 100%. However, the classification accuracy of satin weave is only 95.99%. When the twist coefficient increases to 370, the classification accuracies of twill and satin fabric are 94.73% and 93.11%, respectively. Hence, it can be found that when the twist coefficient of the colored spun yarn increases, the spiral disposition of the dyed fibers is more compact, and the internal and external transfer of the fibers is more intense, resulting in more complex fiber distribution on the fabric surface. Therefore, in order to further study the validity and stability of the texture representation model when the twist coefficient changes, this paper analyzes fabrics with the same blending coefficient of dyed fibers through experiments. The results are shown in Table 9.

Table 9.

Classification results of the first batch of samples with the same quality ratio of dyed fibers

Quality ratio of dyed fibers (%)	Sample number	Twist factor	Classification accuracy (%)	Average classification accuracy (%)
White 30.00 Red 35.00 Green 35.00	SS41A	330	100.00	100.00
	SS43A	370	100.00	100.00
	SS41D	330	100.00	100.00
	SS43D	370	100.00	100.00
	SS41F	330	85.55	92.77
	SS43F	370	100.00	92.77
White 30.00 Red 34.75 Green 35.25	SS51A	330	100.00	100.00
	SS53A	370	100.00	100.00
	SS51D	330	100.00	100.00
	SS53D	370	100.00	100.00
	SS51F	330	100.00	92.97
	SS53F	370	85.95	92.97
White 30.00 Red 34.50 Green 35.50	SS61A	330	100.00	100.00
	SS63A	370	100.00	100.00
	SS61D	330	100.00	86.66
	SS63D	370	73.33	86.66
	SS61F	330	100.00	100.00
	SS63F	370	100.00	100.00
White 30.00 Red 34.00 Green 36.00	SS71A	330	100.00	100.00
	SS73A	370	100.00	100.00
	SS71D	330	100.00	100.00
	SS73D	370	100.00	100.00
	SS71F	330	93.33	96.67
	SS73F	370	100.00	96.67
White 30.00 Red 33.00 Green 37.00	SS81A	330	100.00	100.00
	SS83A	370	100.00	100.00
	SS81D	330	100.00	100.00
	SS83D	370	100.00	100.00
	SS81F	330	93.33	96.67
	SS83F	370	100.00	96.67

The results show that the texture representation model established in this paper can effectively express the texture change caused by the blending differences of dyed fibers and the changes in the twist coefficient of colored spun yarns. The classification accuracy of plain weave is stable at 100%, and those of twill and satin weave are kept above 90%, which verifies the effectiveness and robustness of the texture representation model. However, it is worth noting that some abnormal fluctuations occur in the experimental results. For example, the classification accuracy of SS63D and SS53F is relatively low. By comparing and analyzing the sample images, it is found that similar fluctuations are mainly due to the abnormal aggregation and stacking of dyed fibers in local areas of some samples, as shown in Figure 4.

Figure 4.

Abnormal aggregation of dyed fibers in colored spun woven fabrics.

The types of dyed fibers and the color of dyed fibers are the key factors in the process of fabric texture change. In order to verify the effectiveness and universality of this method, the second batch of 26 samples of colored spun woven fabrics mixed with cotton fibers were analyzed, and the results are shown in Table 10. Similarly, there were not only significant color difference among samples, but also a wider range of quality ratio of dyed fibers, ranging from 0.25% to 8.0%. The samples are only plain weave.

Table 10.

Experimental results of the second batch of colored spun woven fabrics

Sample number	Color of dyed fibers	Quality ratio of dyed fibers (%)	Classification accuracy (%)	Average classification accuracy (%)
17001	White : Red : Yellow	95.05:3.93:1.02	100	100
17002		94.60:3.90:1.48	100
17003		94.10:3.90:2.00	100
17004		93.60:3.90:2.50	100
17005		93.10:3.90:3.00	100
17006		94.96:3.06:1.98	100
17007		94.51:3.52:1.97	100
17008		94.10:3.90:2.02	100
17009		93.38:4.60:2.02	100
17010		92.90:5.10:2.00	100
17011		93.90:4.60:1.50	100
17012		94.00:3.53:2.47	100
17014		93.75:4.70:1.55	100
17015		94.01:3.53:2.46	100
17029		94.30:3.76:1.94	100
17030		94.00:4.00:2.00	100
17031		94.00:4.00:2.00	100
17032		94.00:4.00:2.00	100
17018	White : Black : Blue	92.00:4.00:4.00	100	100
17019		90.00:4.00:6.00	100
17020		88.00:4.00:8.00	100
17021		90.00:2.00:8.00	100
17022		89.00:3.00:8.00	100
17023		88.10:4.00:7.90	100
17024		91.00:3.00:6.00	100
17025		92.00:2.00:6.00	100

The results show that the classification accuracy of plain weave fabrics is stable at 100%. Therefore, the change of the fibrous materials and the color of the dyed fibers have no effect on the classification of the fabric, which is robust.

Comparison results

As the key supporting technology of artificial intelligence, deep learning has been widely used in many fields, with high classification accuracy. Among them, the deep belief network (DBN) and the convolutional neural network (CNN) are relatively mature algorithms for image texture feature extraction.^19,20 In order to fully verify the effectiveness and practicability of this method, a comparative experiment scheme is established. The DBN uses the deep learning toolbox provided by MATLAB and records it as method 1. The CNN uses the Inception V3 model of Google Tensor Flow for transfer learning, retains the convolution layer of the network model and modifies it to a fully connected layer, which records it as method 2. There are 716 training samples of methods 1 and 2; the comparison results are shown in Table 11.

Table 11.

Comparative experimental results

Experimental method	Classification accuracy %			Average classification accuracy (%)
Experimental method	Plain	Twill	Stain	Average classification accuracy (%)
Method 1	100.00	93.33	90.00	94.44
Method 2	96.67	91.67	93.33	93.89
Proposed method	100.00	97.33	95.45	97.59

The experimental results show that, compared with the classical deep learning algorithm of the DBN and CNN, the texture representation model established in this paper can effectively fuse the spatial and frequency domain texture features of the woven fabric image, and has the ability of effective and stable texture description. In the process of fabric texture classification, the overall classification accuracy is 97.59%, which is better than that of the comparison method.

Conclusions

In this paper, a texture representation model of colored spun woven fabric is proposed, which combines spatial and frequency features. This model uses the fusion features of the LBP and LPQ to describe fabric images, and realizes the classification of colored spun woven fabrics by means of a hierarchical hybrid classifier. The experimental results show that the LPQ feature is robust to the fuzzy transformation and the texture representation model has stronger ability of texture description and anti-noise than the single local binary mode feature. At the same time, compared with the deep learning algorithm, the texture representation algorithm can ensure a high classification accuracy.

The research in this paper can provide technical support for the reconstruction, classification and defect detection of colored spun woven fabric images. Moreover, whether fibrous materials and the color of dyed fiber have an impact on the texture change of colored spun fabrics will be the authors’ next research direction.

Footnotes

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors disclosed receipt of the following financial support for the research, authorship and/or publication of this article: This work was supported by the Natural Science Foundation of Hubei Province (No. 2014CFB754), the Young Talents Project of Science and Technology Research of the Hubei Provincial Department of Education (No. Q20141607) and the Science and Technology Guiding Project of the China National Textile and Apparel Council (No. 2018035 and No. 2014072).

ORCID iDs

Li Yuan

Xue Gong

Muli Liu

References

Jing

. Automatic recognition of weave pattern and repeat for yarn-dyed fabric based on KFCM and IDMF. Optik 2015; 216: 2876–2883.

Lin

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

Wang X, Georganas ND and Petriu EM. Automatic woven fabric structure identification by using principal component analysis and fuzzy clustering. In: 2010 IEEE instrumentation and measurement technology conference proceedings, Austin, TX, 3–6 May 2010, paper no. IMTC.2010.5488107, pp.590-595. Piscataway, NJ: IEEE.

Jing JF, Liu SM, Li PF, et al. The fabric defect detection based on CIE L*a*b* color space using 2-D Gabor filter. J Text Inst. 2015; 7: 1–9.

Yuan

Chen

, et al. A defect detection method for unpatterned fabric based on multidirectional binary patterns and the gray-level co-occurrence matrix. Text Res J 2019; 90: 776–796.

Li PF, Zhao Z, Zhang L, et al. The real-time vision system for fabric defect detection with combined approach. In: Zhang YJ (ed) International conference on image and graphics, Tianjin, China, 13-16 August 2015, paper no. LNCS 9219, pp.460–473. Cham: Springer.

Xiao

Wang

, et al. 2D-LBP: an enhanced local binary feature for texture image classification. IEEE Trans Circ Syst Video Technol 2019; 29: 2796–2808.

Guo

Zhang

. A completed modeling of local binary pattern operator for texture classification. IEEE Trans Image Proc 2010; 19: 1657–1663.

Fernandez

Alvarez

Bianconi

. Image classification with binary gradient contours. Opt Laser Eng 2011; 49: 1177–1184.

10.

Pawening RE, Dijaya R, Brian T, et al. Classification of textile image using support vector machine with textural feature. In: 2015 international conference on information and communication technology and systems (ICTS), Surabaya, Indonesia, 16 September 2015, paper no. ICTS.2015.7379883, pp.119–122. Piscataway, NJ: IEEE.

11.

Tajeripour F, Kabir E and Sheikhi A. Defect detection in patterned fabrics using modified local binary patterns. In: international conference on computational intelligence and multimedia applications, Sivakasi, Tamil Nadu, 13–15 December 2007, paper no. ICCIMA.2007.50, pp.261–267. Piscataway, NJ: IEEE.

12.

Liu

Yan

, et al. Fabric defect detection based on sparse representation of main local binary pattern. Int J Clothing Sci Technol 2017; 29: 282–293.

13.

Xue

Cheng

. Intelligent detection of defects of yarn-dyed fabrics by energy-based local binary patterns. Text Res J 2012; 82: 1960–1972.

14.

Zhang

Jing

Zhang

. Fabric defect classification based on LBP and GLCM. J Fiber Bioeng Informatics 2015; 8: 81–89.

15.

Yuan

Dai

. Coloration characteristics and influence factors of colored spun fabric based on primary-color fibers blending. J Text Res 2018; 39: 38–43.

16.

Ojaja

Pietikainen

. Multi-resolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Trans Pattern Anal Mach Intell 2002; 24: 971–987.

17.

Zhang B, Liu GY and Xie GQ. Facial expression recognition using LBP and LPQ based on Gabor wavelet transform. In: 2016 2nd IEEE international conference on computer and communications (ICCC), Chengdu, China, 14–17 October 2016, paper no. CompComm. 2016.7924724, pp.365–369. Piscataway, NJ: IEEE.

18.

Xue

Titterington

. Joint discriminative generative modeling based on statistical tests for classification. Pattern Recognit Lett 2010; 9: 1048–1055.

19.

Fan R and Hu W. Face recognition with improved deep belief networks. In: 2017 13th IEEE international conference on natural computation, fuzzy systems and knowledge discovery (ICNC-FSKD), Guilin, China, 29–31 July 2017, paper no. FSKD.2017.8393043, pp.1822–1826. Piscataway, NJ: IEEE.

20.

Vennelakanti A, Shreya S, Rajendran R, et al. Traffic sign detection and recognition using a CNN ensemble. In: 2019 IEEE international conference on consumer electronics (ICCE), Las Vegas, NV, 11–13 January 2019, paper no. ICCE.2019.8662019, pp.1–4. Piscataway, NJ: IEEE.