A universal defect detection approach for various types of fabrics based on the Elo-rating algorithm of the integral image

Abstract

In order to overcome the shortcoming that a fabric defect detection method can only fit into a certain type of fabric, this paper presents a novel method by integrating the idea of the integral image into the Elo-rating algorithm (IIER), which can detect the defects of various types of fabric speedily. Firstly, the golden sub-blocks are extracted from defect-free images. The whole images are divided into many sub-blocks, and the integral images of these sub-blocks are obtained. Next, the R sub-blocks are randomly selected from these integral sub-blocks, and each block is assigned an initial Elo point. Afterwards, the R sub-blocks are matched against all sub-blocks and the Elo points are updated after each competition. Finally, regions with bright defects accumulate high Elo points and regions with dark defects accumulate low Elo points. Thus, the threshold value image can be obtained by thresholding the final Elo points, in which white, gray and black regions correspond to bright, dark-defect and defect-free regions, respectively. The performance of the proposed method is evaluated on databases of three categories of fabric, namely raw fabric, yarn-dyed fabric and patterned fabric. The experimental results show that the IIER is a universal algorithm, which has high detection rate for different types of fabrics; in particular, the average correct detection rate can reach 100% for dot-patterned fabric. In addition, the detection time can be significantly reduced comparing with the Elo-rating algorithm (ER). Particularly for star-patterned fabric, the average detection time per image is 24.18 seconds less than the ER.

Keywords

defect detection Elo-rating integral image raw fabric yarn-dyed fabric patterned fabric

In the textile industry, fabric defect detection plays an important role in the product quality control process. Due to different kinds of fabric defects, textile enterprises are faced with huge losses. It is estimated that the price of fabrics will fall by 45–65% due to defects.¹ At present, most fabric inspection is carried out visually by human workers, which is not only costly, but is also inaccurate and unreliable due to human error and visual fatigue. In order to reduce the loss and improve the quality, more and more textile enterprises pay attention to automatic fabric defect detection. However, the wide variety of fabrics to be tested, the diverse types of defects and the long detection time have brought great challenges for defect detection.^2,3 The application of computer vision technology in fabric automatic vision detection^4–6 not only provides an efficient, low-cost and accurate detection method to replace the labor force, but also expands the detection ability to cover a wider range of different fabric types.^7–17

In recent years, several fabric defect detection algorithms have been proposed.^18–38 These methods are used to detect potential patterns that are different from non-defective textures by analyzing image textures. Raheja et al.,¹⁸ Miao,¹⁹ Bi et al.²⁰ and Zhu et al.²¹ used the local mean and standard deviation of gray texture features or the gray co-occurrence matrix to detect fabric defects. Firstly, it is necessary to determine the image grayscale distribution. Then through the gray level-gradient co-occurrence matrix, the change rate of the image grayscale is calculated. According to the change rate of the image grayscale, it can judge whether there are defects or not. Although good detection results have been achieved, there is still room for improvement in the calculation and identification. Jing et al.,^22,23 Jing,²⁴ Bissi et al.²⁵ and Wang et al.²⁶ used a Gabor filter to filter and extract the texture features of the defect texture, and used golden image subtraction or the principal component analysis (PCA) algorithm to segment the defect from the normal texture. Good results have been achieved in the detection of raw fabric, but the versatility of the algorithm needs to be further improved. Li et al.²⁷ and Ngan et al.²⁸ processed the sample image using the wavelet decomposition algorithm and reconstructed the new image using wavelet coefficients. The defect can be detected by subtracting the new image from the original image. The algorithm has a good effect on the detection of raw fabric and patterned fabric, but the detection of yarn-dyed fabric still needs further research. Cohen et al.,²⁹ Hua et al.³⁰ and Yang³¹ used the Markov random field model to transform the image texture into state space. In some cases, the estimated model parameters make the Markov covariance uncertain, resulting in detection failure. At the same time, choosing appropriate model parameters is also a problem. In addition, the computational amount of the algorithm is large and the real-time performance is poor. Leng et al.,³² Serdatoglu et al.,³³ Kraoananthavaram et al.³⁴ and Anitha and Radha³⁵ used the independent component analysis algorithm to extract the texture and structure information from images. The structure information can be reduced by phase consistency. Then the template image can be distinguished from the input image features. Wang et al.³⁶ and Zhang et al.³⁷ treat a defective region as a salient region in fabric images. A novel fabric defect detection method, which is based on the saliency metric for color dissimilarity and positional aggregation, is proposed. In the method, the RGB color space of a given fabric image is first converted into the CIE L*a*b color space for feature representation. Then, the color dissimilarity and the positional distance between similar patches are jointly used to measure the defective values. The experimental results of several fabric images show that the proposed method has certain effectiveness. Tsang et al.³⁸ pioneered a novel fabric defect detection method based on the Elo-rating algorithm (ER). In the ER method, an image is divided into standard size sub-blocks with a start-up reference point. The Elo points of these sub-blocks are updated through matches between different sub-blocks. Any sub-blocks with defects will be shown in the resultant threshold value image: a white resultant image corresponds to a bright defect and a gray resultant image corresponds to a dark defect. The ER is a promising method for defect detection. However, the detection time is long, the efficiency is low and the defect detection of the algorithm in raw fabric and yarn-dyed fabric is not mentioned in the literature.

The existing methods in the above literature are effective in detecting defects of a certain type of fabric, but cannot be used to detect other types of defects, that is, they have poor universality. In view of this, this paper proposes a universal defect detection method based on the Elo-rating algorithm of the integral image (IIER). The advantages of the employed method are as follows.

By extending the ER to multiple types of fabric defect detection, including raw fabric, yarn-dyed fabric and patterned fabric, the versatility of the defect detection algorithm is improved. Through experiments, the selection method of two parameters, namely the number of randomly selected sub-blocks and the size of the sub-blocks, is summarized when the IIER is used in different types of fabric defect detection.

In the ER, repeated competitions are performed between sub-blocks, and each match is compared pixel by pixel. The amount of calculation is very large, and the detection time is very long. The introduction of integral image overcomes the fatal shortcoming of the traditional ER.

The IIER used in this paper has achieved good detection results in different types of fabric defect detection, and the highest total average detection rate can reach 95%. Particularly in the detection of dot-patterned fabric, the average detection rate can reach 100%. In addition, the rapidity of defect detection is greatly improved by the algorithm. Particularly in the defect detection of star-patterned fabric, the average detection time is 24.18 seconds less than the ER.

The structure of this paper is organized as follows. In the second section, the traditional ER is introduced and its time-complexity is analyzed. In addition, the introduced integral image is described. In the third section, the proposed approach is described and its time is analyzed. Moreover, the influence of the use of integral image on the detection time is also introduced. In the fourth section, the influence of three parameters in the proposed algorithm on the experimental results is summarized. In the fifth section, the detection results and analysis of the experiment are provided to validate our method on three databases of real fabric images. Finally, the sixth section gives the conclusion.

Traditional Elo-rating algorithm

The ER was first proposed by Professor Elo,^39–42 and is widely used in international chess, football, basketball and other sports. Tsang et al.³⁸ extended this algorithm to fabric defect detection for the first time. For the ER, an image is divided into many sub-blocks and each sub-block is considered as a player. There are two parts in the ER: the threshold value computation section and detection section. The threshold value computation section mainly calculates two threshold values. One is used to judge the win or loss of a sub-block match, and the other is used to judge whether there is defect or not in the sub-block. The execution process of the detection section is the same as that of the threshold value computation section.

The threshold value computation is as follows.

The golden sub-block Q with size of $m \times n$ is extracted from a defect-free image with size $M^{'} \times N^{'}$ pixels. According to the size of the golden sub-block Q, the defect-free image can be divided into $(M' - m + 1) \times (N^{'} - n + 1)$ sub-blocks of the size $m \times n$ .

Slide Q on each pixel along each row on the defect-free image and all the sub-blocks on it are matched with Q. The score s of the match against the corresponding position is calculated after each move of the sliding process.

For any two sub-blocks $Q_{A}$ and $Q_{B}$ with size of $m \times n$ , the score s of sub-block $Q_{A}$ is defined as Equation (1) after matching against sub-block $Q_{B}$

s = \frac{1}{mn} \sum_{i = 1}^{m} \sum_{j = 1}^{n} (Q_{A} - Q_{B})

(1)

In Equation (1), $Q_{A}$ and $Q_{B}$ represent the pixels at the corresponding positions of any two sub-blocks of size $m \times n$ . The integers $i (i = 1, 2, \cdot \cdot \cdot, m)$ and $j (j = 1, 2, \cdot \cdot \cdot, n)$ represent the position coordinates.

After the golden sub-block Q completes the matches with the whole defect-free image, a score matrix $S = {s}$ of size $(M^{'} - m + 1) \times (N^{'} - n + 1)$ is outputted according to the corresponding score of all sub-blocks. Then maximum $S_{max}$ and minimum $S_{min}$ of S are extracted.

Repeat the above steps for k defect-free images and obtain k maximum $S_{max}$ and minimum $S_{min}$ . The average values of k maximum $S_{max}$ and minimum $S_{min}$ are calculated respectively to obtain the threshold values of $T_{win}$ and $T_{lose}$ , which are used to determine the win or loss of a sub-block match.

Set the initial Elo point of $(M^{'} - m + 1) \times (N^{'} - n + 1)$ sub-blocks as 1000.

The R sub-blocks are randomly selected from the divided $(M^{'} - m + 1) \times (N^{'} - n + 1)$ sub-blocks. The whole $(M^{'} - m + 1) \times (N^{'} - n + 1)$ sub-blocks will have R matches against the sub-blocks R randomly selected. After each match, the Elo point of the corresponding position on the Elo point matrix will be updated.

After step 5 has finished, the updated Elo point matrix $E^{*}$ is obtained. Then the maximum $E_{max}^{*}$ and minimum $E_{min}^{*}$ of $E^{*}$ are extracted.

Repeat the above steps for k defect-free images (the parameter k refers to the number of defect-free images that need to be trained) and obtain k maximum $E_{max}^{*}$ and minimum $E_{min}^{*}$ . The average values of k maximum $E_{max}^{*}$ and minimum $E_{min}^{*}$ are calculated respectively to obtain the threshold values of $T_{bright}$ and $T_{dark}$ , which are used to determine whether there is a defect or not in sub-block.

Detection

Carry out the same operation as the above steps 4–6 of the threshold value computation section in the tested sample to obtain an updated Elo point matrix $E^{*}$ .

According to the updated Elo point matrix $E^{*}$ , the threshold values $T_{bright}$ and $T_{dark}$ are obtained, and then the final three-color (black–gray–white) resultant image is obtained.

Time-complexity analysis of the traditional ER

The time-complexity of the ER is mainly reflected in the computational complexity of the algorithm. We take C2R2 raw fabric as an example for analysis.

Computational analysis of the threshold value section

The threshold value computation section is operated on defect-free images. It is assumed that the size of the defect-free image is 64 × 64, and the size of the golden sub-block is 7 × 4. For each sample, it can be divided into (64 – 7 + 1) × (64 – 4 + 1) = 58 × 61 = 3538 sub-blocks of size 7 × 4. Then the golden sub-block Q is used to slide the pixel by pixel on the defect-free image, and is matched against all sub-blocks in the defect-free image. For the match of any two sub-blocks, 7 × 4 = 28 operations are performed according to Equation (1). For a golden sub-block Q to compete with the 3538 sub-blocks of the whole image, 28 × 3538 = 99,064 operations are performed. In the ER, five defect-free images are selected for training, and a total of 5 × 99,064 = 495,320 operations are performed. Then, the R sub-blocks are randomly selected from 3538 sub-blocks, and are matched against the whole 3538 sub-blocks. The values of R are different for different fabric types. For C2R2 raw fabric, the value of R is 10. The R sub-blocks are matched against the other sub-blocks in five images, and a total of 10 × 5 × 28 × 3538 = 4,953,200 operations are performed. The more complex the texture feature and structure feature of the fabric, the larger the value of R will be, and the greater the computing times will be. Therefore, the total number of calculations in the threshold value section is 495,320 + 4,953,200 = 5,448,520 in the traditional ER.

Computational analysis of the detection section

The calculation of the detection section is similar to the threshold value computation section, but is operated on tested images. Therefore, the total number of calculations in the detection section is 10 × 28 × 3538 = 990,640 in the traditional ER.

Based on the above analysis, it can be seen that the calculation of the traditional ER is very considerable due to a great amount of comparison, especially if it is used in defect detection. For fabrics with a complex texture, structure and cycle, the calculation is larger. Therefore, we need to find a method to reduce the calculation amount of the ER. In the ER, when the golden sub-block Q or the random sub-block R slides pixel by pixel on the whole image, the operation is to first calculate the pixel difference of each point, and then sum, as shown in the previous equation (1). Deform Equation (1) to get Equation (2)

\begin{matrix} s = \frac{1}{mn} \sum_{i = 1}^{m} \sum_{j = 1}^{n} (Q_{A} - Q_{B}) \\ = \frac{1}{mn} (\sum_{i = 1}^{m} \sum_{j = 1}^{n} Q_{A} - \sum_{i = 1}^{m} \sum_{j = 1}^{n} Q_{B}) \end{matrix}

(2)

Each of the terms $\sum_{i = 1}^{m} \sum_{j = 1}^{n} Q_{A}$ or $\sum_{i = 1}^{m} \sum_{j = 1}^{n} Q_{B}$ in the transformed operation above is the principle of the integral image algorithm. Therefore, we add the integral image on the traditional ER to see if it can reduce the time of detection.

Integral image

The simple representation of the integral image is that the pixel value of each point is equal to the sum of all pixel values before the point, which can be expressed by the following equation

I_{Σ} (x, y) = \sum_{i = 1}^{x} \sum_{j = 1}^{y} I (i, j)

(3)

In Equation (3), $I_{Σ} (x, y)$ represents the integral image of point $(x, y)$ , and $I (i, j)$ represents the pixel of the corresponding position on the image. The integers $i (i = 1, 2, \cdot \cdot \cdot, x)$ and $j (j = 1, 2, \cdot \cdot \cdot, y)$ represent the position coordinates. If the pixel sum of a region in the image is calculated, the integral image of the end point of the region can be used for the calculation.

In Figure 1, the pixel sum of the region Q can be obtained by the integral images of points A, B, C and D, which can be expressed by the following equation

I_{Q} = I_{A} - I_{B} - I_{C} + I_{D}

(4)

Figure 1.

Integral image of the region Q.

In Equation (4), $I_{A}$ , $I_{B}$ , $I_{C}$ and $I_{D}$ represent integral images of points A, B, C and D, respectively.

Elo-rating algorithm of the integral image

The proposed method, the IIER, is a new perspective for the detection of fabric defects. In the IIER method, the number of accumulated Elo points is used to judge bright defective regions, dark defective regions and defect-free regions of an image. The bright defects and dark defects for different types of fabric samples are shown in Figure 2. Among them, the first, second and third columns are defect-free samples, bright defect samples and dark defect samples, respectively.

Figure 2.

The bright defects and dark defects for different types of fabric samples: (first column) defect-free samples; (second column) defective samples with bright defects; (third column) defective samples with dark defects.

In the long run, the Elo points accumulated is a fair indication of the player's performance. In fabric defect detection, sub-blocks of a fabric image are considered as the players and a match is regarded as the matrix operation between these sub-blocks. The IIER method shown in Figure 3 also involves two phases: the threshold value computation section and the detection section. The threshold value computation section is operated on defect-free images, which contain two threshold values. The first threshold value is used to determine the win or loss of a sub-block match. The second threshold value is used to determine whether there is defect or not in the sub-block. The detection section is performed on the tested images, and the execution process is the same as the threshold value computation section, while each sub-block is judged as to whether there is defect by the second threshold value obtained in the threshold value computation section.

Figure 3.

Flow of the Elo-rating algorithm of the integral image.

Threshold value computation

The first threshold value is used to judge the win or loss of each sub-block, and as the basis of the Elo points updating of each sub-block. The second threshold value is used to judge whether there is defect or not in each sub-block. It is the key to accurately detecting the defect. The specific process of the section is as follows.

Perform a level 2 Haar wavelet transform on defect-free images to obtain corresponding approximated images of size $M \times N$ pixels. A level 2 Haar wavelet transform is used because it reduces the computational amount and the running time (including training time and testing time) in subsequent stages. The running time of different level Haar wavelet transforms is shown in Table 1. Level 2 strikes a balance between image size and computational time. The approximated image of a level 3 Haar wavelet would be too coarse for fabric inspection, so it is not considered. All the following steps are operated on the approximated images of a level 2 Haar wavelet transform.

The golden sub-block Q with size of

m \times n

is extracted from each approximated image according to the type of fabric, densities, texture characteristics, organizational structure and the thickness of the yarn. Different yarns, fabric densities and organizational structure are shown in different fabrics with different fabric textures. When the texture features are not obvious or the yarn is thick, the size of Q is chosen to be larger. This will be a better confirmation of defects.

Table 1.

Running time of different level Haar wavelet transforms

		Without wavelet	Level 1 Haar wavelet	Level 2 Haar wavelet
	Image size (pixels)	256 × 256	128 × 128	64 × 64
Running time (s)	Training time	168	41.22	9.58
Running time (s)	Testing time	27.38	9.94	2

Then get the integral image $I_{Q Σ}$ of the golden sub-block Q as Equation (5)

I_{Q Σ} = \sum_{i = 1}^{m} \sum_{j = 1}^{n} I_{Q} (i, j)

(5)

where

I_{Q} (i, j)

represents the pixel of the corresponding position on the image and the integers

i (i = 1, 2, \cdot \cdot \cdot, m)

and

j (j = 1, 2, \cdot \cdot \cdot, n)

represent the position coordinates.

The integral image of the entire approximate image is constructed. According to the size $m \times n$ of the golden sub-block Q, the approximate image can be divided into $(M - m + 1) \times (N - n + 1)$ sub-blocks of the same size as the golden sub-block. The integral image of each sub-block in the approximate image is obtained according to Equation (4).

Slide $I_{Q Σ}$ on the whole integral image and the integral image of all sub-blocks are matched against $I_{Q Σ}$ . The score s of the match corresponding position is calculated after each move of the sliding process.

For the integral images of any two sub-blocks $I_{A Σ}$ and $I_{B Σ}$ with size of $m \times n$ , the scores s of integral sub-block $I_{A Σ}$ are defined as in Equation (6) after matching against integral sub-block $I_{B Σ}$

s = \frac{1}{mn} (I_{A Σ} - I_{B Σ})

(6)

After $I_{Q Σ}$ completes the matches with the whole integral image, a score matrix $S = {s}$ of size $(M - m + 1) \times (N - n + 1)$ is outputted according to the corresponding score of all sub-blocks

S = \begin{matrix} {\begin{matrix} s_{1, 1} & s_{1, 2} & \dots & s_{1, (N - n + 1)} \\ s_{2, 1} & s_{2, 2} & \dots & s_{2, (N - n + 1)} \\ : \\ s_{(M - m + 1), 1} & s_{(M - m + 1), 2} & \dots & s_{(M - m + 1), (N - n + 1)} \end{matrix}} \end{matrix}

(7)

Then the maximum $S_{max}$ and minimum $S_{min}$ of S are extracted.

Repeat the above steps 1–4 for k (k = 5) defect-free images. Then the first threshold value judging the win or loss of each sub-block can be obtained as Equation (8)

T_{win} = \bar{S_{k_{max}}} T_{lose} = \bar{S_{k_{min}}}

(8)

where

\bar{S_{k_{max}}}

and

\bar{S_{k_{min}}}

are the average values of

S_{max}

and

S_{min}

corresponding to five defect-free images, respectively.

According to the first threshold values $T_{win}$ and $T_{los s}$ , the result (win/tie/loss) of each match is determined. The sub-block A is used as an example. If the scores s of the sub-block A are greater than $T_{win}$ , it wins the match, whose actual score $X_{a}$ is 1. If the scores s of the sub-block A are less than $T_{los s}$ , it loses the match, whose actual score $X_{a}$ is 0. If the scores s of the sub-block A are greater than $T_{los s}$ and less than $T_{win}$ , it ties the match, and the actual score $X_{a}$ is 0.5

\begin{matrix} {\begin{matrix} if s > T_{win}, & A is defined as a winner, & X_{a} = 1 \\ if T_{loss} < s < T_{win}, & A is defined as a tie & X_{a} = 0.5 \\ if s < T_{loss}, & A is defined as a losser, & X_{a} = 0 \end{matrix} \end{matrix}

Assign the initial Elo point of each sub-block as 1000. Then a reference matrix of Elo points $E = {1000}_{(M - m + 1) (N - n + 1)}$ is obtained for future updates.

Assume that the initial Elo points of sub-blocks A and B are $E_{a}$ and $E_{b}$ , respectively. The expected values of sub-blocks A and B winning the match are $Ex p_{a}$ and $Ex p_{b}$ , respectively. They are defined as in Equations (9) and (10), respectively

Ex p_{a} = \frac{1}{1 + 10^{\frac{E_{b} - E_{a}}{400}}}

(9)

Ex p_{b} = 1 - Ex p_{a}

(10)

The R sub-blocks are randomly selected from the divided $(M - m + 1) \times (N - n + 1)$ integral sub-blocks. The $(M - m + 1) \times (N - n + 1)$ sub-blocks of whole integral image will have R matches against the R integral sub-blocks, respectively. After each match, the Elo point of the corresponding position on the Elo point matrix will be updated. We can repeat step 7 for the $(M - m + 1) \times (N - n + 1)$ sub-blocks.

The updating formula is as Equation (11), according to the position of sub-block

E^{*} = E_{a} + K (X_{a} - Ex p_{a})

(11)

where

E^{*}

is the updated Elo point matrix of the corresponding position,

E_{a}

is the initial Elo point, K is a constant and

Ex p_{a}

is the expected value of the corresponding sub-block winning the match.

X_{a}

is the actual match score determined by step 5 and its value can be 1, 0.5 or 0.

After step 7 has finished, the updated Elo point matrix $E^{*}$ is obtained

E^{*} = \begin{matrix} {\begin{matrix} E_{1, 1} & E_{1, 2} & \dots & E_{1, (N - n + 1)} \\ E_{2, 1} & E_{2, 2} & \dots & E_{2, (N - n + 1)} \\ : \\ E_{(M - m + 1), 1} & E_{(M - m + 1), 2} & \dots & E_{(M - m + 1), (N - n + 1)} \end{matrix}} \end{matrix}

(12)

Then the maximum

E_{max}^{*}

and minimum

E_{min}^{*}

E^{*}

are extracted.

Repeat the above steps 6–8 for k (k = 5) defect-free images. (The parameter k refers to the number of defect-free images that need to be trained.) Then the second threshold values judging defect can be obtained as Equation (13)

T_{bright} = \bar{E_{k_{max}}^{*}} T_{dark} = \bar{E_{k_{min}}^{*}}

(13)

where

\bar{E_{k_{max}}^{*}}

and

\bar{E_{k_{min}}^{*}}

are the average values of

E_{max}^{*}

and

E_{min}^{*}

corresponding to five defect-free images, respectively.

Detection

The detection section is implemented on the basis of the threshold value computation section, and it is only carried out on the tested sample. After obtaining the updated matrix of the tested sample, the second threshold values are used for threshold value segmentation. The specific process of the section is as follows.

Perform a level 2 Haar wavelet transform on the tested image to obtain the corresponding approximated image of size $M \times N$ pixels.

The integral image of the whole approximated tested image is constructed. The integral image is divided into a number of sub-blocks with the size $m \times n$ , and the integral image of each sub-block is obtained.

Carry out the same operation as the above steps 6–8 of the threshold value computation section in the tested sample to obtain an updated Elo point matrix $E^{*}$ .

According to the updated Elo point matrix $E^{*}$ and the threshold values $T_{bright}$ and $T_{dark}$ judging defect, the final three-color (black–gray–white) resultant image U is obtained

\begin{matrix} {\begin{matrix} if E^{*} > T_{bright}, & U = 1, sub-block is bright defect \\ if T_{dark} < E^{*} < T_{bright}, & U = 0, sub-block is defect-free \\ if E^{*} < T_{dark}, & U = 0.5, sub-block is dark defect \end{matrix} \end{matrix}

Consequently, the size of the final detection result image U is $(M - m + 1) \times (N - n + 1)$ . It is outputted after median filtering. The bright defect will be displayed as white, the dark defect as gray and defect-free as black.

Time analysis of the IIER

In order to more accurately compare the computational amount, this part uses the same fabric as above, namely C2R2 raw fabric.

Computational analysis of the threshold value section

The approximate image size of C2R2 raw fabric is 64 × 64, and a golden sub-block Q of size 7 × 4 is extracted. For each sample, it can be divided into (64 – 7 + 1) × (64 – 4 + 1) = 58 × 61 = 3538 sub-blocks of size 7 × 4. Firstly, to calculate the integral image of the golden sub-block Q, it is necessary to perform 7 × 4 – 1 = 27 calculations. Secondly, to obtain the integral image of the whole approximate image, it is necessary to perform 64 × 64 – 1 = 4095 calculations. Thirdly, the integral image of the 3538 sub-blocks according to the size of the golden sub-block Q is obtained in the whole integral image, for which it is necessary to perform (64 – 7 + 1) × (64 – 4 + 1) – 1 = 3537 calculations. Finally, the golden sub-block Q is matched against all sub-blocks of the whole integral image, for which it is necessary to perform (64 – 7 + 1) × (64 – 4 + 1) = 3538 calculations. Therefore, for an image, starting from the extraction of the golden sub-block Q to the completion of the competition, it is necessary to perform 27 + 4095 + 3537 + 3538 = 11,197 calculations. As with the ER, we still select five images for training, so a total of 5 × 11,197 = 55,985 calculations are required. Since the training image of the second threshold is the same as that of the first threshold, the corresponding integral image can be directly applied to the second threshold value. The R sub-blocks randomly selected on the integral image are matched against all the sub-blocks in it. If R is 10, 10 × (64 – 7 + 1) × (64 – 4 + 1) = 35,380 calculations are performed. Finally, in five images of training, the total number of calculations is 5 × 35,380 = 176,900. Therefore, the total number of calculations in the threshold value section is 55,985 + 176,900 = 232,885 in the IIER.

Computational analysis of the detection section

The calculation of the detection section is similar to that of the threshold value computation section. Firstly, the integral image of the tested image is obtained. Secondly, the integral image of the corresponding sub-block in the whole image is obtained according to the size of the golden sub-block. Finally, the R sub-blocks randomly selected from the whole integral image are matched against other sub-blocks in it. The total number of calculations in the detection section is 4095 + 3537 + 35,380 = 43,012 in the IIER.

Table 2 shows the comparison of the computational amount between two algorithms. Through the analysis of the computational amount by the traditional ER and IIER, it can be seen that the IIER can significantly reduce the computational amount. In the threshold section, the time can be reduced by 95.7%, and in the detection section, the time can be reduced by 95.6%. The theoretical basis is provided for the application of this method to the online defect detection of fabric.

Table 2.

The comparison of the computational amount (operand) between two algorithms

	The computational amount of threshold value section		The computational amount of detection section
	ER	IIER	ER	IIER
Computational amount	5,448,520	232,885	990,640	43,012
Difference	5,215,635		947,628
Percentage of time reduction	95.7%		95.6%

ER: Elo-rating algorithm; IIER: Elo-rating algorithm of the integral image.

The effect of the integral image on time

The use of the integral image in the IIER converts the pixel-by-pixel operation used in the ER into an image block operation. In the third section, through the algorithm analysis, it can be seen that the idea of the integral image is added into the ER, which greatly reduces the computational amount. The reduced computational amount means that the computational time will be reduced and the speed of defect detection will be improved. Table 3 shows the average detection time of the ER and the IIER and their differences for different types of fabric defect detection. In Table 3, 20 samples are taken for each type of fabric and are tested by two algorithms. The average detection time of each algorithm is obtained.

Table 3.

The average detection time of the Elo-rating algorithm (ER) and the Elo-rating algorithm of the integral image (IIER)

	C1R1	C1R3	C2R2	C2R3	Yarn-dyed fabric	Dot-	Star-	Box-
Average detection time of ER (s)	4.54	9.61	10.65	23.34	4.62	17.62	31.13	21.47
Average detection time of IIER (s)	3.03	3.61	3.23	4.64	3.03	5.14	6.95	5.66
Differences of time (s)	1.51	6.0	7.42	18.7	1.59	12.48	24.18	15.81

It can be seen from Table 3 that the average detection time of the two algorithms is significantly different. Both computation complexity and the average detection time all prove the superiority of the IIER. Since the IIER is compared and detected in the form of image blocks, and the operations of these blocks are independent of each other, the proposed algorithm has a promising result to meet real-time requirements in actual production by adopting parallel operations.

Parameter analysis

In the IIER, the selection of several parameters plays a key role in detecting defects accurately. The parameters are as follows: constant K, the number of sub-blocks R randomly selected and the size of sub-block $m \times n$ .

The selection of constant K

A greater K indicates that the closer match has more significant influence. Generally, K is in the range of 10–32, and the value of K is shown in Table 4. In the paper, the influence of K that changed the detection results is shown in Figure 4. The experiment proved that the value of K has no significant influence on the detection results. In some types of image detection, a smaller value of K will make the defect detection process incomplete. However, overall, the higher the value of K is, the better the detection effect is. Therefore, in this paper, K takes its maximum value of 30 in the allowable value.

Figure 4.

The influence of K on the detection results: (first column) defective images; (second column) constant K = 10; (third column) constant K = 20; (fourth column) constant K = 30.

Table 4.

The corresponding value of constant K

Elo points	Value of K
Elo points ≤ 2000	30
2000 < Elo points < 2400	20
Elo points ≥ 2400	10

The selection of the number of sub-blocks R randomly selected

In the case of the certain size m × n of sub-blocks, the running time and detection success rate (DSR) of the algorithm are affected directly by the number of sub-blocks R randomly selected. Running time includes the training time used in the threshold value computation and the testing time used in defect detection. When the size of the sub-blocks is $10 \times 8$ , $7 \times 4$ , $4 \times 4$ and $2 \times 2$ , respectively, the influence of the number of sub-blocks R randomly selected on the running time and the DSR is as shown in Figure 5. In Figure 5, 50 images of each size are taken for experiment, and the average running time and DSR of 50 images of each size were calculated respectively.

Figure 5.

The influence of the number of sub-blocks R randomly selected on the running time and detection success rate.

As shown in Figure 5(a), with the increase in the number of sub-blocks R randomly selected, the running time of the images of each size increases accordingly. When R is greater than 50, the running time increases rapidly, which has a great impact on the real-time detection. As shown in Figure 5(b), when the number of sub-blocks R randomly selected is equal to five, the DSR of each sub-block image is very low. However, when R is greater than or equal to 10 but less than 30, images of each size show a high DSR. When R is greater than 40, the DSR of images of each size is significantly reduced. To sum up, in order to achieve a high DSR and a short running time, it is appropriate to select the number of sub-blocks R randomly between 10 and 30.

The number of sub-blocks R randomly selected for different types of fabrics is different. It is related to the texture, density and thickness of the yarn. However, it can be seen from previous experiments that the number of sub-blocks R randomly selected should be selected between 10 and 30 to meet the dual requirements of the running time and DSR. In Figures 6–8, we take five cases of R as 5, 10, 15, 20 and 30, respectively, to analyze the influence of the number of sub-blocks R on the detection results of three different types of fabric samples. The optimal parameter is the one with a high DSR and low running time.

Figure 6.

The influence of the number of sub-blocks R randomly selected on the detection results of raw fabric.

Figure 7.

The influence of the number of sub-blocks R randomly selected on the detection results of yarn-dyed fabric.

Figure 8.

The influence of the number of sub-blocks R randomly selected on the detection results of patterned fabric.

As can be seen from Figures 6–8, if the size of the sub-blocks of different types of fabrics is set to a fixed value, the number of randomly selected sub-blocks R will have a great impact on the detection results, and the defect area will increase with the increase of the number of sub-blocks R randomly selected. If the value of R is too small, the training process will have poor applicability and the outline of the detected defects will probably be incomplete. On the contrary, if the value of R is too large, the training process will be time-consuming, which affects the speed of detection. In addition, the detected defect area will be much larger than the actual defect, and the detection results cannot really reflect the real characteristics of the contour and shape of the reactive defect.

In general, the accuracy of the test is measured by the following parameters: (a) DSR; (b) True Positive Rate (TPR); (c) True Negative Rate (TNR); and (d) False Positive Rate (FPR). The four parameters above are represented by True Negative (TN), False Negative (FN), False Positive (FP) and True Positive (TP). In Table 5, TP, FP, TN and FN in defect detection are defined.

Table 5.

Definitions of TP, FP, TN, FN in defect detection

	Actually defect-free	Actually defective
Detected as defect-free	True negative (TN)	False Negative (FN)
Detected as defective	False positive (FP)	True Positive (TP)

The DSR, TPR, TNR, FPR can be defined as Equations (14)–(17), respectively

DSR = \frac{TN + TP}{TN + FP + TP + FN}

(14)

TPR = \frac{TP}{TP + FN}

(15)

TNR = \frac{TN}{TN + FP}

(16)

FPR = \frac{FP}{FP + TN}

(17)

In order to analyze the influence of the number of sub-blocks R randomly selected on the DSR, the C2R2 raw fabric is used as an example to analyze the relationship between the number of sub-blocks R randomly selected and the DSR, TPR, TNR and FPR. The reason for using the C2R2 raw fabric as an example is that the four types of defects included in the C2R2 raw fabric have the same parameter selection in the detection. The results are easy to compare.

On the basis of sub-block size

m \times n = 7 \times 4

, the values of R in the experiments were 10, 15, 20, 30 and 40, respectively. The numerical results of the experiments are shown in Table 6. As can be seen from Table 6, compared with other values of R, when R is 10, the DSR, TPR and TNR are all the largest, while the FPR is the smallest. It is indicated that 10 is the optimal number of sub-blocks randomly selected for fabric defect detection. This proves that the earlier selection of parameters is reasonable.

Table 6.

The numerical results between the detection success rate (DSR), True Positive Rate (TPR), False Positive Rate (FPR) and True Negative Rate (TNR) versus randomly selected sub-blocks R

The value of R	The value of TN, TP, FN and FP				DSR (%)	TPR (%)	FPR (%)	TNR (%)
The value of R	TN	TP	FN	FP	DSR (%)	TPR (%)	FPR (%)	TNR (%)
R = 10	190	190	9	1	97.4	95.5	0.5	99.5
R = 15	165	165	11	24	90.4	93.8	12.7	87.3
R = 20	162	162	11	27	89.5	93.6	14.3	85.7
R = 30	156	156	11	33	87.6	93.4	17.5	82.5
R = 40	153	153	11	36	86.7	93.3	19	81

TN: True Negative; TP: True Positive; FN: False Negative; FP: False Positive.

The relationship between the DSR, TPR, FPR and TNR versus randomly selected sub-blocks R is shown in Figure 9. It can be seen from Figure 9 that when R is 10, the DSR, TPR and TNR are all particularly high. This shows that the detection effect is the best, which further proves that the parameters are selected properly.

Figure 9.

The relationship between the detection success rate (DSR), True Positive Rate (TPR), False Positive Rate (FPR) and True Negative Rate (TNR) versus randomly selected sub-blocks R.

The selection of the size $m \times n$ of sub-blocks

Because different fabrics have different texture characteristics, for different fabrics and different defects, the selection of the size $m \times n$ of sub-blocks has a direct impact on the accurate detection and detection time under the premise that the number of sub-blocks selected randomly is determined. In the IIER, the R sub-blocks are randomly selected from the divided $(M - m + 1) \times (N - n + 1)$ integral sub-blocks. The integral sub-blocks are segmented according to the size of the golden sub-block Q. Therefore, the randomly selected sub-block R and the golden sub-block Q have the same size. The size has a significant influence on whether the algorithm can detect defects accurately.

In Figures 10–12, on the premise of a certain number of randomly selected sub-blocks R, we take five cases of $m \times n$ as $2 \times 2$ , $4 \times 4$ , $7 \times 4$ , $6 \times 6$ , $10 \times 6$ , respectively, to analyze the influence of the size of the sub-blocks on the detection results (including the DSR and detection time) obtained from the experiments of three different types of fabric samples. The optimal parameter is the one with a high DSR and low running time.

Figure 10.

The influence of the size of sub-blocks on the detection results of raw fabric. DSR: detection success rate.

Figure 11.

The influence of the size of sub-blocks on the detection results of yarn-dyed fabric. DSR: detection success rate.

Figure 12.

The influence of the size of sub-blocks on the detection results of patterned fabric. DSR: detection success rate.

As can be seen from Figures 10–12, compared with the size of the texture unit of the sample itself, the sub-block size is too large or small to obtain good detection results. If sub-block size selection is not appropriate, not only can it not correctly reflect the original shape of the defect, but also it increases the probability of missed or false detection. Therefore, the sizes of the texture unit of different fabric samples determine the optimal size of the sub-blocks. Therefore, the selection of different fabric sub-block sizes is closely related to the size of the texture unit, density and yarn size of the fabric sample. The smaller the density, the finer the yarn, and the smaller the sub-block will be. Because small sub-blocks are enough to show the fine features of the fabric, if larger sub-blocks are selected, there is the possibility of false detection. On the contrary, the higher the density and the thicker the yarn, the larger the sub-block will be. Too small a sub-block cannot fully reflect the characteristics of the fabric, resulting in the possibility of missing inspection.

Figure 13 illustrates the DSR, TPR, FPR and TNR plots of the effect of the sub-block size $m \times n$ on the dot-patterned, box-patterned and star-patterned fabrics. The three types of fabrics used the same evaluation processes with the sub-block sizes of $2 \times 2$ , $3 \times 2$ , $6 \times 6$ , $7 \times 4$ and $10 \times 6$ . Figure 13(a) clearly indicates that the $7 \times 4$ size provides the best TPR, DSR and TNR for the dot-patterned fabric among the five choices, the $6 \times 6$ size provides the best TPR, DSR and TNR for the box-patterned fabric in Figure 13(b) and the $10 \times 6$ size provides the best TPR, DSR and TNR for the star-patterned fabric in Figure 13(c). This further proves that the previous parameter selection is completely reasonable.

Figure 13.

The relationship between the detection success rate (DSR), True Positive Rate (TPR), False Positive Rate (FPR) and True Negative Rate (TNR) versus the size of sub-blocks for (a) dot-patterned fabrics, (b) box-patterned fabrics and (c) star-patterned fabrics.

For the box-patterned fabrics, when $m \times n$ is $2 \times 2$ and $3 \times 2$ , respectively, the FPR and TNR are not calculated because both the FP and TN are zero. The same as for the star-patterned fabrics, when $m \times n$ is $2 \times 2$ , the FPR and TNR are not calculated because both FP and TN are zero.

The experimental results show that the IIER proposed in this paper can not only detect the defect accurately, but also enhance the versatility greatly. In particular, the detection method greatly shortens the detection time and improves the rapidity of detection. The algorithm is effective for detecting raw fabric, and also has a good detection effect in yarn-dyed fabric and patterned fabric. It improves the disadvantage of poor versatility of the defect detection algorithm, and has very important research value.

Detection results and analysis

In this paper, the IIER method is applied to detect raw fabric, yarn-dyed fabric and patterned fabric. In this paper, the detected fabric images contain 1545 samples, with 1280 defective and 265 defect-free samples. Among them, the raw fabric images contain 1000 samples, with 800 defective and 200 defect-free samples. The yarn-dyed fabric images contain 450 samples, with 400 defective and 50 defect-free samples. The patterned fabric images contain 95 samples, with 80 defective and 15 defect-free samples. The sample size of the three types of fabric images is 256 × 256 pixels. The raw fabrics contain four types of defects (namely Hole, Oil Spots, Thread Errors, Objects on the Surface). The yarn-dyed fabrics contain several kinds of common defects (such as Belt Yarn, Knots, Thick Bar, Broken Warp, Broken Pick, Oil Stain and so on). The patterned fabrics contain six types of defects (namely Broken End, Hole, Netting Multiple, Thick Bar, Thin Bar, Knots).

The defect detection of raw fabric

In the experiment, the algorithm is used to detect the defect samples in the TILDA fabric sample database. The TILDA fabric sample database is a standard fabric defect sample database, which contains two groups of fabric samples, C1 and C2. Fabric group C1 contains fabric images of uniform texture with fine yarns and thin texture, so their fabric density is high, such as woven fabrics. Fabric group C2 contains fabric images of visible texture with thick yarns and grid-like structures; thus, their fabric density is small, such as knitted fabrics. Fabric groups C1 and C2 were composed of 400 defective samples and 100 defect-free samples, respectively. The defect types of C1 and C2 include Holes (E1), Oil Spots (E2), Thread Errors (E3) and Objects on the Surface (E4). The defect detection results of raw fabric are shown in Figure 14.

Figure 14.

The defect detection results of raw fabric.

Since the yarn of group C2 is thicker and less dense than the yarn of group C1, it is not easy to display defects in the smaller sub-block. Therefore, the size of the gold sub-block Q of group C2 is larger than that of group C1. Due to different yarns and different fabric textures in the same group, the choice of parameters is not the same. Through experiments, we can find that the two defects of Holes and Oil Spots are more obvious than the two defects of Thread Errors and Objects on the surface of the fabric. Therefore, in the same texture fabric, the size of the sub-blocks required to detect the two defect types of Holes and Oil Spots is smaller than that of the other two defect types, and the number of sub-blocks randomly selected is also less than or equal to the number of sub-blocks of the other two defect types. However, the C2R3 fabric has thicker yarn and Oil Spots are not obvious on it, so the sub-block size of the fabric is relatively large. Specific parameters and test results are shown in Table 7. Through the detection of different types of defects, the average detection rates of C1R1, C1R3, C2R2 and C2R3 are 93.5%, 97%, 95% and 94.5%, respectively. The total average detection rate of raw fabric is 95%.

Table 7.

Defect detection rate of two groups, C1 and C2

Fabric type	Defective name	Parameter setting (R, m, n, K)	Detection rate (%)	Average detection Rate (%)	Total average detection rate (%)
C1R1	Holes	(15, 2, 2, 30)	94	93.5	95
	Oil Spots	(15, 2, 2, 30)	98
	Thread Errors	(20, 4, 4, 30)	92
	Objects on the Surface	(20, 4, 4, 30)	90
C1R3	Holes	(15, 2, 2, 30)	96	97
	Oil Spots	(15, 2, 2, 30)	94
	Thread Errors	(15, 4, 4, 30)	100
	Objects on the Surface	(15, 4, 4, 30)	98
C2R2	Holes	(10, 7, 4, 30)	98	95
	Oil Spots		100
	Thread errors		90
	Objects on the Surface		92
C2R3	Holes	(10, 7, 4, 30)	100	94.5
	Oil Spots	(10, 10, 8, 30)	98
	Thread Errors	(20, 7, 4, 30)	88
	Objects on the Surface	(20, 7, 4, 30)	92

The defect detection of yarn-dyed fabric

In order to test the effectiveness of the algorithm, yarn-dyed fabric is used as the second type of detection objects. Fabric images with 10 different colors are detected, respectively, including dark-red, white-blue, navy-blue, orange, white, pink, gray-white, gray-black and so on. Since the yarns of the yarn-dyed fabric are relatively thin, yarn-dyed fabrics of different colors are close relatively in thickness, and the yarn density is also uniform. Therefore, in the use of the IIER to detect fabric defects, their parameters are selected consistently. The sub-block size $m \times n$ is set as 2 × 2, the number of sub-blocks R selected randomly is 15 and constant K is 30. The defect detection results of yarn-dyed fabric are shown in Figure 15. Among them, the first, third and fifth columns show the defective images, while the second, fourth and sixth columns show the detection results.

Figure 15.

The defect detection results of yarn-dyed fabric: (first, third and fifth columns) defective images; (second, fourth and sixth columns) detection results.

The defect detection results of yarn-dyed fabric are shown in Table 8. In Table 8, 10 colors are taken in yarn-dyed fabric, each yarn-dyed fabric contains five types of defects and eight samples are detected for each type of defect. Therefore, 40 samples are detected for each yarn-dyed fabric. According to the statistical results of Table 8, it can be seen that the average detection rates of the 10 types of yarn-dyed fabric are 97.5%, 95%, 95%, 97.5%, 87.5%, 90%, 82.5%, 95%, 90% and 82.5%, respectively. The total average detection rate of yarn-dyed fabric can reach 91.3%. The DSR of the common defect is higher in the industrial field, such as Belt Yarn, Knots, Oil Stain, Holes, Corruption and so on. However, due to the relatively complex background texture of the fabric, the defect and background texture are difficult to distinguish, resulting in a relatively low success rate of individual fabric defect detection, which is the focus of our later research and improvement.

Table 8.

Defect detection rate (%) of yarn-dyed fabric

Yarn-dyedfabric Defect type	Dark-red	White-blue	Navy-blue	Orange	White	Pink	Black-gray	Gray-white	Gray-black	Blue-white
Belt yarn	100	100	100	100	75	87.5	87.5	100	100	75
Knots	100							87.5	100	87.5
Thick Bar			100	100				87.5	75
Broken Warp	87.5				75	100	75
Broken Pick	100	87.5	100				75	100	75	75
Oil Stain		100		87.5	100	100	75
Holes		100		100	87.5	87.5	100		100	100
Corruption		87.5	75	100	100			100
Neps	100		100			75				75
Average detection rate	97.5	95	95	97.5	87.5	90	82.5	95	90	82.5
Total average detection rate	91.3

The defect detection of patterned fabric

Through the detection of two different types of fabric samples, it is proved that the IIER has good detection performance. In order to verify the versatility of the algorithm better, patterned fabric is selected as the third kind of detection object. The detected samples consist of three types of fabric: dot-patterned, star-patterned and box-patterned. Among them, star-patterned and box-patterned fabrics contain five kinds of defect types: Broken End, Hole, Netting Multiple, Thick Bar and Thin Bar. In addition to the five kinds of defect types above, dot-patterned fabric also includes the defect of Knots, a total of six kinds of defect types. Each kind of defect type includes five tested samples, so there is a total of 80 samples.

The repeated pattern is made up of repeated cells, and the patterned fabric is formed by translation and rotation of these cells. Considering the particularity of patterned fabric texture, the effective size of the sub-blocks is set based on the texture features of three different patterns of fabric and the size of their repeated cells in our experiments. The defect detection results of patterned fabric are shown in Figure 16. Among them, the first column shows the defective names, the second, fourth and sixth columns show the defective images and the third, fifth and seventh columns show the detection results.

Figure 16.

The defect detection results of patterned fabric: (first column) defective names; (second, fourth and sixth columns) defective images; (third, fifth and seventh columns) detection results.

In Table 9, the detection parameters and the results of patterned fabric are given; the average detection rates of the three patterned fabrics are 100%, 92% and 88%, respectively. According to the data in the table, 75 of the 80 defective fabric samples are successfully detected, with the total average detection rate of 93.3%.

Table 9.

Defect detection rate (%) of patterned fabric

Fabric type	Parameter setting (R, m, n, K)	Broken End	Hole	Netting Multiple	Thick Bar	Thin Bar	Knots	Average detection rate	Total average detection rate
Dot-	(20, 7, 4, 30)	100	100	100	100	100	100	100
star-	(20, 10, 6, 30)	100	80	100	100	80	–	92	93.3
box-	(20, 6, 6, 30)	100	80	60	100	100	–	88

Through the statistical analysis of patterned fabric defect detection, the IIER has achieved ideal results in patterned fabric defect detection, which further proves that the algorithm has strong versatility and application value.

Results analysis and practical application

Results analysis

In this section, we compare the detection results of the proposed algorithm in this paper with those of other algorithms. In order to make a fair comparison, we only use the defect images employed by other algorithms for comparison. The results of the proposed method compared with other methods are shown in Figure 17.

Figure 17.

The defect detection results analysis: (first column) defective sample names; (second column) defective images; (third column) detection results of the Frangi filtering method; (fourth column) detection results of the Expectation Maximization (EM) method; (fifth column) detection results of the Gabor method; (sixth column) detection results of the Vector Quantized Principal Component Analysis (VQPCA) method; (seventh column) detection results of the Elo-rating algorithm (ER) method; (eighth column) detection results of the Elo-rating algorithm of the integral image (IIER) method.

As can be seen from Figure 17, for the raw fabric, both the Frangi filter combined with the fuzzy C-means algorithm and the Expectation Maximization (EM) algorithm can detect defects accurately, but the outline of the defects detected are incomplete and even some noise may exist. In contrast, the IIER proposed in this paper is more prominent in the details of defects.

For yarn-dyed fabric, the Gabor filter is a commonly used method to detect the defects. However, its detection results are not satisfactory, especially for fabrics whose defects are not so obvious. The IIER is very good at defect detection.

For patterned fabric, we compared Vector Quantized Principal Component Analysis (VQPCA), the traditional ER and the IIER. It can be seen from the detection results that there is much noise in the detection process of the VQPCA algorithm. The ER and the IIER have similar detection results. However, the IIER is much better at showing the details of the defects. In addition, in the previous section we also mentioned that the IIER takes less detection time than the ER.

In general, the proposed IIER has achieved good detection results for each type of fabric defect, while other algorithms have their respective strengths and shortcomings; this proves the proposed algorithm has a strong universality.

Practical application

As can be seen from the above description, all three types of fabrics have achieved good detection results by only modifying the parameters of the IIER, which means that the proposed algorithm can be used in actual applications. For the convenience of applying the algorithm in a real textile industrial system, we further design a pseudo-code as shown in Algorithm 1 to illustrate the process of implementation for defect detection.

Algorithm 1.

The detection process of the IIER

Require: take k defect-free images in textile industrial system.

1: for each defect-free image

2: according to the steps 1–4 of threshold value computation in the third section

the image is preprocessed

the golden sub-blocks extracted are matched against the whole image.

3: end

4: output: the threshold value judging the win or loss of each sub-block according to step 5

5: take k the initial Elo point matrix about to the k defect-free images

6: for each defect-free image

7: R sub-blocks randomly selected to match against the whole image

8: update the Elo point matrix according to the calculation method in steps 7 and 8

9: end

10: output: the second threshold values judging defect according to step 9

11: take a defect image

12: perform steps 6–8 on it

13: output: an updated Elo point matrix

14: judge defect according to the second threshold

15: output: the final detection result image

Conclusion

A new algorithm for the automatic detection of fabric defects named the IIER is proposed in this paper, in which the defect detection is similar to carrying out fair matches. The algorithm is helpful to improve the quality control of fabric products. The addition of the integral image provides the possibility to improve the speed of defect detection. The algorithm also depends on the setting of two parameters: the number of sub-blocks R adaptive selected and the size $m \times n$ of sub-blocks. Both of these parameters are closely related to the texture characteristics, structure features and the thickness of yarn. The proposed algorithm shows good versatility and rapidity in the test results, so it has wide application value in actual fabric defect detection. In the future, additional theoretical development involving game theory for matches as it relates to defect detection should be carried out. Such research will contribute to extend the IIER to more defect detection in the textile, glass, ceramics, tile and printed circuit board industries.

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 Key Project “Heterogeneous feature structure fusion and modelling for human action recognition” of the Natural Science Foundation of Shanxi Province of China (Grant No. 2017JZ020) and the Science and Technology Project “Automatic defect detection for web offsetting based on machine vision” of the Xi’an Bureau, Shanxi, China (Grant No. 2017080CG/RC043 (XALG020)).

References

Qian W. Research on fabric defect detection based on neural network. Donghua University, 2018.

Ngan

HYT

Pang

GKH

Yung

NHC

. Automated fabric defect detection—a review. Image Vis Comput 2011; 29: 442–458.

Ngan

HYT

Yuan

, et al. Patterned fabric inspection and visualization by the method of image decomposition. IEEE Trans Autom Sci Eng 2014; 11: 943–947.

Mera

Orozco-Alzate

Branch

, et al. Automatic visual inspection: an approach with multi-instance learning. Comput Ind 2016; 83: 46–54.

Kumar

. Computer-vision-based fabric defect detection: a survey. IEEE Trans Ind Electron 2008; 55: 348–363.

Mahajan

Kolhe SR and Patil

. A review of automatic fabric defect detection techniques. Adv Comput Res 2009; 1: 18–29.

Jing

Yang

. Defect detection on patterned fabrics using distance matching function and regular band. J Eng Fiber Fabric 2015; 10: 85–97.

Ngan

HYT

Pang

GKH

. Regularity analysis for patterned texture inspection. IEEE Trans Autom Sci Eng 2009; 6: 131–144.

Jing

Fan

, et al. Yarn-dyed fabric defect detection based on deep-convolutional neural network. J Text Res 2017; 38: 68–74.

10.

Jing

Fan

, et al. Yarn-dyed fabric defect detection based on Gaussian back substitution image decomposition. J Text Res 2016; 37: 136–141.

11.

Zhang

Chen

. Deep neural network for halftone image classification based on sparse auto-encoder. Eng Appl Artif Intell 2016; 50: 245–255.

12.

Chen S, Feng J and Zou L. Study of fabric defects detection through Gabor filter based on scale transformation. In: Lam D.K.T. (eds) International conference on image analysis and signal processing IEEE, Zhejiang, China, 9–11 April 2010, pp.97–99.

13.

Karlekar VV, Biradar MS and Bhangale KB. Fabric defect detection using wavelet filter. In: Hans Robert T. (eds) International conference on computing communication control and automation IEEE, Langkawi, Malaysia, 2-4 September 2014 , pp.712–715.

14.

Zhang

Chen

, et al. Sparsity-based inverse halftoning via semi-coupled multi-dictionary learning and structural clustering. Eng Appl Artif Intell 2018; 72: 43–53.

15.

Zhang

Duan

. Color inverse halftoning method with the correlation of multi-color components based on extreme learning machine. Appl Sci 2019; 9: 841.

16.

RZh

Zhang

Jing

, et al. Fabric defect detection based on Gaussian mixture models of EM algorithm. J Comput Eng Appl 2014; 50: 184–187.

17.

Zhang

RZh

Jing

, et al. Fabric defect detection based on Frangi filter and fuzzy C-means algorithm in combination. J Text Res 2015; 36: 120–124.

18.

Raheja

Kumar

Chaudhary

. Fabric defect detection based on GLCM and Gabor filter: a comparison. Optik 2013; 124: 6469–6474.

19.

Miao

. Image texture detection based on parallel gray level grade co-occurrence matrix. Laser Infrared 2011; 41: 1287–1291.

20.

Bi MD, Zou C, Sun ZHG, et al. Real-time textural defect detection based on label run length co-occurrence matrix. In: Dai Wenting, Li Yongju, Jin Weidong (eds) IEEE international conference on transportation, ChangChun, China,16-18 December 2011, pp.271–274.

21.

Zhu DD, Pan RR, Gao WD, et al. Yarn-dyed fabric defect detection based on autocorrelation function and GLCM. Autex Res J 2015; 15: 226–232.

22.

Jing

Yang

, et al. Supervised defect detection on textile fabrics via optimal Gabor filter. J Ind Text 2014; 44: 40–57.

23.

Jing

Fan

. Automated fabric defect detection based on multiple Gabor filters and KPCA. Int J Multimedia Ubiquitous Eng 2016; 11: 93–106.

24.

Jing

. Automatic defect detection of patterned fabric via combining the optimal Gabor filter and golden image subtraction. J Fiber Bioeng Informatic 2015; 8: 229–239.

25.

Bissi

Baruffa

Placidi

, et al. Automated defect detection in uniform and structured fabrics using Gabor filters and PCA. J Vis Commun Image Represent 2013; 24: 838–845.

26.

Wang

QCh

Jing

Zhang

, et al. Denim defect detection based on optimal Gabor filter. J Laser Optoelectron Progr 2018; 55: 363–369.

27.

Zhang

Jing

, et al. Fabric defect detection based on multi-scale wavelet transform and Gaussian mixture model method. J Text Inst Proc Abs 2015; 106: 587–592.

28.

Ngan

HYT

Pang

GKH

Yung

, et al. Wavelet based methods on patterned fabric defect detection. Pattern Recognit 2005; 38: 559–576.

29.

Cohen

Fan

Attali

. Automated inspection of textile fabrics using textural models. IEEE Trans Pattern Anal Mach Intell 1991; 13: 803–808.

30.

Hua

Huang

. Fabric defect detection using GMRF model. J Donghua Univ Eng Ed 1999; 16: 10–13.

31.

Yang

. Fabric defect detection of statistic aberration feature based on GMRF model. J Text Res 2013; 34: 137–142.

32.

Leng

Zhang

Fan

, et al. Fabric defect detection using independent component analysis and phase congruency. Wuhan Univ J Nat Sci 2014; 19: 328–334.

33.

Serdatoglu A, Ertuzin A and Erci A. Textile fabric defect detection with wavelet and independent component analysis. In: Amasyali M.F. (eds) IEEE signal processing & communications applications conference, Kayseri, Turkey, 16–18 May 2005, pp.103–106.

34.

Kraoananthavaram

Rao

Krishna Prasad

. Automatic defect detection of patterned fabric by using RB method and independent component analysis. Int J Comput Appl 2012; 39: 52–56.

35.

Anitha S and Radha V. Evaluation of defect detection in textile images using Gabor wavelet based independent component analysis and vector quantized principal component analysis. In: Mohan S, Kumar SS (eds) Proceedings of the fourth international conference on signal and image processing, New Delhi, India, 13–15 December 2012, pp.433–442.

36.

Wang

ChT

, et al. Application of algorithm with improved frequency-tuned salient region in fabric defect detection. J Text Res 2018; 39: 154–160.

37.

Zhang

Yan

, et al. Fabric defect detection using salience metric for color dissimilarity and positional aggregation. IEEE Access 2018; 6: 49170–49181.

38.

Tsang

CSC

Ngan

HYT

Pang

GKH

. Fabric inspection based on the Elo rating method. Pattern Recognit 2016; 51: 378–394.

39.

Ouyang MS. Research on athletics evaluation system based on ELO rating. Wuhan: Wuhan University of Technology, 2013.

40.

Pelánek

. Applications of the Elo rating system in adaptive educational systems. Comput Educ 2016; 98: 169–179.

41.

Albers

PCH

Han

. Elo-rating as a tool in the sequential estimation of dominance strength. Animal Behav 2001; 61: 489–495.

42.

Elo AE. The Rating of Chess Players, Past and Present. New York: Arco, 1978.

A universal defect detection approach for various types of fabrics based on the Elo-rating algorithm of the integral image

Abstract

Keywords

Traditional Elo-rating algorithm

Detection

Time-complexity analysis of the traditional ER

Computational analysis of the threshold value section

Computational analysis of the detection section

Integral image

Elo-rating algorithm of the integral image

Threshold value computation

Detection

Time analysis of the IIER

Computational analysis of the threshold value section

Computational analysis of the detection section

The effect of the integral image on time

Parameter analysis

The selection of constant K

The selection of the number of sub-blocks R randomly selected

The selection of the size m × n of sub-blocks

Detection results and analysis

The defect detection of raw fabric

The defect detection of yarn-dyed fabric

The defect detection of patterned fabric

Results analysis and practical application

Results analysis

Practical application

Conclusion

Footnotes

Declaration of conflicting interests

Funding

References

The selection of the size $m \times n$ of sub-blocks