FABRIC YARN DETECTION BASED ON

IMPROVED FAST R-CNN MODEL

Haiyan Xu*

Business School, Yangzhou Polytechnic Institute, Yangzhou, Jiangsu, 225127, China

hyxu_ypi@163.com

Reception: 16/11/2022 Acceptance: 09/01/2023 Publication: 06/03/2023

Suggested citation:

X., Haiyan (2023). Fabric yarn detection based on improved fast R-CNN

model. 3C TIC. Cuadernos de desarrollo aplicados a las TIC, 12(1), 287-306.

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

287

ABSTRACT

With the rapid development of modern computer technology, and gradually combined

with the textile industry, the application of modern computer technology in the field of

textile is increasingly extensive, which makes textile production gradually move

towards the road of automation development. This paper proposes an automatic

detection method of simple weave fabric density based on computer image vision.

Computer vision and digital image processing technology are used to analyze and

identify the simple weave fabric's warp and weft yarn information and calculate the

fabric density. To avoid the phenomenon of warp and weft yarn skew, a method of

fabric skew correction based on the Radon transform is proposed. The optimal

decomposition order of these four fabrics is k = 2, k = 5, and k = 3. The decomposition

series is k. It is found that the relative error of both warp and weft density is about

1.00%. Most of the data obtained by the method of correlation coefficient curve to

determine the optimal decomposition series are consistent with the results of the

energy curve method. The relative error of the density test results of No. 3 fabric, No.

6 fabric, and No. 7 fabric is higher than 10%, and the relative error of No. 3 fabric is

the highest, reaching 66%. This shows serious errors in these three fabrics' warp and

weft density. To solve the problems of simple weave fabric density detection, the

corresponding algorithm is used to solve the problems. Finally, good results are

obtained, which verifies the feasibility of this method. It is significant to realize the

automatic measurement of fabric density in textile factories.

KEYWORDS

Fabric yarn; Fast R-CNN algorithm; Visual inspection; Image processing; Wavelet

decomposition.

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

288

PAPER INDEX

ABSTRACT

KEYWORDS

1. INTRODUCTION

2. IMPROVED FAST R-CNN ALGORITHM

2.1. Basic structure and characteristics of convolution neural network

2.2. R-CNN series algorithm

2.3. Yolo algorithm

3. SIMPLE WEAVE FABRIC DENSITY DETECTION BASED ON WAVELET

TRANSFORM

3.1. Wavelet transform

3.2. Wavelet transform of simple weave fabric

3.3. Determination of optimal wavelet decomposition series of fabric image

4. SIMPLE WEAVE FABRIC DENSITY TEST RESULTS AND ANALYSIS

4.1. Calculation of warp and weft density of simple weave fabric processed by

computer

4.2. Analysis of experimental results

5. CONCLUSION

DATA AVAILABILITY

CONFLICT OF INTEREST

REFERENCES

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

289

1. INTRODUCTION

With the rapid development of modern computer technology, and gradually

combined with the textile industry, the application of modern computer technology in

the field of textile is increasingly extensive, which makes textile production gradually

move towards the road of automation development. The structure parameters such as

fabric texture, density, and color yarn arrangement are important in detecting and

controlling textile quality. Currently, most factories and enterprises in the textile

industry mainly rely on manual sample analysis and detection of fabric texture with the

aid of fabric lenses, which is subjective, time-consuming, labor-intensive, and prone to

errors. Therefore, the use of computer image processing technology to effectively

replace manual to achieve intelligent fabric density detection, improve industrial

production efficiency, achieve automation, and intelligent production of textile products

is of great significance.

In recent years, the rapid development of computers has made computer vision

technology into people's field of vision and has been much attention. Under our

research and exploration, computer vision continues to develop and progress, and

image processing technology has been widely used. Especially in textile testing,

computer vision technology is also used, which makes the textile industry more

intelligent and efficient. Shukla et al. Collected the transmission image and reflection

image of fabric sample by optical principle, calculated the autocorrelation value of

each row and column of the image with the autocorrelation function after

preprocessing, and processed and analyzed the transmission image and reflection

image respectively to obtain the relevant information of fabric texture parameters [1].

Finally, the fabric structure is determined by the length and weft of each row, and the

fabric structure is determined by scanning the length and weft of each row[2]. Raj et

al. Reduced the gray image level through histogram equalization, then constructed the

gray level co-occurrence matrix according to the pixel spacing and angle changes,

calculated its eigenvalue, and obtained the fabric density parameter through its

periodic calculation [3]. Used autocorrelation function to determine the position,

density, and weave point position of fabric warp and weft yarn. Later, the data of the

organization point area was input into the neural network to output the fabric structure,

which was trained repeatedly. Finally, the fabric structure was identified by a neural

network [4]. Wu and Cao used the gray projection method to get the gray projection

curve in the warp and weft direction of the fabric image[5,6]. The warp and weft yarns

were separated according to the position and quantity of the gray projection curve

peak and valley, and the fabric warp and weft density was calculated. Trafton et al.

First calculated the weft density of twill and satin fabrics was by the gray projection

method and then calculated the warp density of twill and satin fabrics was through the

relationship between the density of twill and satin fabrics and the fabric texture. Later,

the study found that fabric warp and weft yarn inclination was easy to occur when

image processing was used to detect the density of fabric warp and weft [7-9].

Monfared and XZ et al. Proposed using Hough transform to obtain the fabric tilt angle,

then making a gray projection on the fabric along the tilt direction, and finally judging

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

290

the yarn gap according to the wave crest of the projection curve to calculate the fabric

warp and weft density [10,11].

Qin uses MATLAB language to carry out a series of preprocessing for woven fabric

images, then carry out wavelet decomposition and reconstruction to separate and

extract the information of warp and weft yarn, and then carry out binarization and

smoothing processing to obtain the distribution image of warp and weft yarn. Finally,

the warp and weft density of the woven fabric is obtained through program calculation

[12-15]. Shi et al. Carried out multi-layer wavelet decomposition on woven fabric

image through wavelet transform, reconstructed single-layer signal and calculated

average brightness value of warp and weft yarn direction image, and finally calculated

warp and weft density according to periodic change of brightness signal [16,17].

Combined image processing technology with time-frequency transform theory,

transformed woven image from the time domain to frequency domain through Fourier

transform, selected characteristic region to filter and separate single group of warp

and weft yarn images. Finally, the adaptive threshold method was used to locate yarn,

count the number of warp and weft yarn, and calculate woven fabric's warp and weft

density [18-20]. Obtained the frequency spectrum of the fabric through a two-

dimensional fast Fourier transform and calculated the fabric warp and weft density

through the correlation between the characteristic change of the frequency spectrum

and the fabric warp and weft density [21]. Using the Halcon algorithm library and

machine vision technology, Niu processed the fabric image by Fourier transform,

analyzed it by Gabor transform, and finally calculated the fabric warp and weft density

through wavelet transform results [22]. Barreto and Shi use Fourier transform and

wavelet transforms to process fabric images, analyze spectrum characteristics,

process interference information, and then transform them into spatial domain through

inverse transformation. Finally, fabric warp and weft density can be calculated by

spatial domain detection or correlation between frequency spectrum features and

warp and weft yarn density [23,24]. Le obtains the power spectrum of woven fabric

image by Fourier transform, then processes the threshold value and calculates the

fabric warp and weft density by using the relationship between the spatial domain and

frequency domain. Secondly, it uses wavelet transform to separate the warp and weft

sub-images of woven fabric. After processing, the ideal warp and weft yarn density

information is obtained. Finally, a computer program automatically calculates the

fabric warp and weft density. Others use the wavelet transform to reconstruct the

fabric spatial domain image according to the spectrum characteristics of the fabric

image to detect the fabric warp and weft density [26,27]. Some also use the deep

learning method to train many samples to obtain stable automatic detection of fabric

density after Fourier transform or wavelet transform processing of fabric [28].

To sum up, there have been a lot of research and achievements in the automatic

detection of fabric warp and weft density, but there are still some problems that have

not been solved perfectly, and so far, it has not been well applied to actual industrial

production. This paper proposes a warp and weft yarn detection method based on the

improved fast R-CNN algorithm, which is of great significance to realize the

automation and intelligent production of textile products.

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

291

2. IMPROVED FAST R-CNN ALGORITHM

The current deep learning algorithm uses a convolutional neural network to

recognize yarn features, and when the number of training samples of the network is

enough, its recognition accuracy and robustness are better than the traditional image

processing technology, so it has a good application prospect.

2.1. BASIC STRUCTURE AND CHARACTERISTICS OF

CONVOLUTION NEURAL NETWORK

A convolutional neural network (CNN) is a feedforward neural network with a depth

structure. It can complete the local sampling and image-sharing task using a neural

network. Since the convolutional neural networks can collect the spatial and channel

information of feature graphs simultaneously, it is mostly used in the backbone

network of the algorithm to achieve feature extraction tasks[29-30]. CNN structure is

generally divided into the following layers: convolution, pooling, activation, and full

connection layer. The algorithm's core is to adaptively and update the convolution

kernel parameters iteratively through the computer's automatic learning, so the

convolution layer's calculation is particularly important. The main purpose of the

convolution layer is to extract features of different receptive fields by convolution

kernels of various sizes.

CNN differs from the common neural network in that the neurons in the previous

layer are only partially connected with the current layer. This connection structure

greatly reduces the number of branches in the network. A convolution kernel is a local

region. Let the convolution kernel go through the whole characteristic graph.

Weight sharing means that a single parameter controls multiple connections

without considering the position relationship of input data. A convolution kernel with

fixed internal weight parameters is used to process the whole graph by a convolution

operation. The convolution kernel is equivalent to the weight of the traditional network.

Each neuron of the traditional network has different weights, but the same set of

convolution kernels is used in feature processing, so the parameters of the

convolution neural network are shared.

Pooling is an important downsampling operation. Its principle is to do some simple

operations on the neurons in the convolution layer through the local correlation and

take the results as the input values of the neurons in the pooling layer. This operation

not only reduces the amount of calculation but also retains valuable information.

Common pooling operations include maximum pooling and average pooling. We can

select different pooling techniques according to the actual situation to prevent model

overfitting and improve network robustness.

The full connectivity layer (FC) is located at the end of the network hierarchy. Its

function is to connect the feature maps of the previous layer output and map all the

features distributed in the front layer network to the output sample space to reduce the

influence of the target location on classification accuracy. The full connection layer can

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

292

be completed in the actual network by convolution, and all feature expressions can be

mapped to one output value by a convolution operation.

2.2. R-CNN SERIES ALGORITHM

In the development of deep learning, R-CNN is the first industrial-level target

recognition and detection algorithm, which has become an important research

direction in the target recognition and detection field. Fast R-CNN, fast R-CNN, and

other algorithms are based on R-CNN, aiming at the shortcomings of the previous

generation of algorithms for continuous optimization expansion research. Through

training and calculation, the R-CNN algorithm filters the targets in the candidate

region, filters out the invalid feature regions and then completes the corresponding

classification according to the task requirements. Compared with the one-stage target

recognition and detection algorithm, the recognition error rate and miss recognition

rate of the series of algorithms are relatively low, but the recognition speed is relatively

slow. We can choose different algorithms for different target recognition and detection

requirements in practical applications.

It overcomes the limitation of traditional machine learning methods. The biggest

contribution of the Alex net network is the introduction of the activation function of the

modified linear unit (re Lu). The introduction of the re Lu activation function can not

only effectively prevent the overfitting phenomenon but also shorten the training

period due to the reduction of calculation.

The target classification method of the R-CNN algorithm is to use SVM to classify

the extracted features and then use the non-maximum suppression algorithm (NMS)

to evaluate the feature area. The high-score region is identified as the target area, and

the overlapping, redundant area is removed to obtain the region with the highest

possibility of the target. An important factor that affects the performance of the target

recognition and detection model is whether the object can be located accurately.

Because the inaccuracy of the target candidate frame region will cause the overlap

area error, it is necessary to modify and regress the candidate frame and then

generate the final prediction frame.

The fast R-CNN algorithm is an efficient target recognition and detection algorithm

based on the R-CNN algorithm and uses a deep neural network. Because of the

shortcomings R-CNN algorithm, the fast R-CNN algorithm has made corresponding

improvements. The fast algorithm uses the output of the intermediate convolution

layer of the vgg-16 network.

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

293

Figure 1. Structure of vgg-16 network model

The fast R-CNN algorithm can complete the task of feature extraction, detection frame

regression, and classification at the same time. The efficiency of this algorithm is far

higher than that of other algorithms of the R-CNN series. The fast R-CNN algorithm

does not need distributed training and testing. Firstly, it proposes a series of anchors

with preset sizes through the regional recommendation network (RPN), then adjusts

the anchor's size several times through the training network and outputs the final

target detection regression box. Since this paper will optimize the fast R-CNN

algorithm, the principle of the algorithm is analyzed in detail from three aspects:

network architecture, RPN structure, and algorithm loss function.

(1)

2.3. YOLO ALGORITHM

Yolo (you only look once) is a target recognition and detection framework based on

suggestion region. Yolo algorithm is different from the two-stage recognition and

detection idea. Yolo algorithm first divides the given image into S×S cells, which cell is

responsible for detecting the center of each target. Compared with the two-stage

detection algorithm, S×S cells are equivalent to the target region of interest, so there

is no need to generate candidate regions through the network similar to RP, and the

detection task can be completed in one step.

(2)

Next, the C conditional category probabilities are predicted by the segmented grid,

and the redundant boundary boxes are removed by non-maximum suppression

(NMS) to get the best result

{ }{ }

( ) ( ) ( )

* * *

1 1

, ,

i j cls i i i reg i i

cls reg

L p t L p p p L t t

N N

λ= +

∑ ∑

i truth

j pred

C pr( object )*IOU=

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

294

(3)

(4)

Yolo is a simple and fast end-to-end algorithm, but the accuracy of Yolo recognition

and detection is not high, and the object positioning is not accurate enough. It is not

good enough for small targets and dense target recognition and detection.

Table 1. Comparison of advantages and disadvantages of target recognition algorithm

3. SIMPLE WEAVE FABRIC DENSITY DETECTION

BASED ON WAVELET TRANSFORM

Wavelet transform is the inheritance and development of the traditional Fourier

transform. Wavelet transform has good adaptability for time-frequency windows.

Different from Fourier transform, the window size cannot change with the change in

frequency. Wavelet can be used for multi-scale analysis, feature extraction, and

analysis of the object's high-frequency and low-frequency information, a new image

( ) ( )

/( )

pr class object

pr class object pr object

( ) ( )

/ * ( )* *

truth truth

i prad i prrd

pr class object pr object IOU pr class object IOU=⋅

Algorithm name advantage shortcoming

R-CNN

1. CNN is proposed to extract

features

2. Map on Pascal VOC

increased from 35.1% to

53.7%

1. The training process is

divided into stages, and the

recognition speed is slow

2. Consumes disk space

Fast R-CNN

1. Using the full winder

network, the

ROI pooling, all feature maps

can be predicted only once

2. The problem of image

distortion and redundant

computation is reduced

1. The method of extracting

candidate regions is

computationally expensive

and repetitive

2. The end-to-end training is

not implemented

Faster R-CNN

1. Propose RPN network

2. Real end-to-end detection

model

3. Recognition accuracy and

speed have been greatly

improved

1. ROI pooling operation

results in precision loss

YOLOv3

1. Faster speed

2. End to end model

1. There is the deviation in the

accuracy of object position

recognition

2. Low recall rate

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

295

processing method. This study uses the wavelet transform to detect simple weave

fabric images' warp and weft density.

3.1. WAVELET TRANSFORM

Compared with the Fourier transform, the wavelet transform has better time-

frequency window characteristics, which has attracted many experts and scholars to

study. In the past ten or twenty years, wavelet transform has developed rapidly and

widely in many scientific and technological fields. Wavelet transform decomposes the

signal into a series of wavelets by scaling and shifting the original wavelet. Compared

with Fourier transform, the wavelet transform overcomes its three shortcomings: first,

Fourier coefficients cannot change with frequency, but wavelet coefficients can; Two

wavelet transforms can well reflect the signal frequency change with time; Wavelet

transform - Fourier transform can solve the problem of variable window size. Wavelet

transforms mainly include continuous wavelet transform and discrete wavelet

transform.

(5)

(6)

Where k is the decomposition series, and t is the displacement Ψ

(x) length. The

basic wavelet function can be obtained Ψ

(x) based on the wavelet transform. The

wavelet sequence function after displacement and decomposition is as follows:

(7)

When the decomposition series and the displacement length

are continuous

variables, the above wavelet transform process is called continuous wavelet transform

(CWT). There is another form of discrete wavelet transform (DWT). In many cases,

the decomposition series K and the displacement length t need to be discretized by

the power series.

(8)

The discrete wavelet sequence function is as follows:

(9)

(x)

C dx

+∞

−∞

Ψʹ

= < ∞

∫

( , ) ( ), ( ) ( )

f k t R

x t

W k t f x x f x dx

−

⎛ ⎞

=Ψ=Ψ⎜ ⎟

⎝ ⎠

∫

( ) , , ; 0

k t

x t

x k t R k

−

⎛ ⎞

Ψ=Ψ ∈ >

⎜ ⎟

⎝ ⎠

( , ) ( ), ( ) ( )

f m n m

x nt

W m n f x x f x dx

⎛ ⎞

−

=Ψ=Ψ⎜ ⎟

⎝ ⎠

∫

( )

m n m

x nt

⎛ ⎞

−

Ψ=Ψ⎜ ⎟

⎝ ⎠

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

296

3.2. WAVELET TRANSFORM OF SIMPLE WEAVE FABRIC

The fabric image can be decomposed into approximate and detailed image

information by one-dimensional wavelet transform on two-dimensional images, low-

frequency, and high-frequency parts. Then the high-frequency part can continue to be

decomposed into a group of high-frequency and low-frequency components, and the

low-frequency part can be further decomposed into another group of high-frequency

and low-frequency components. Finally, a two-dimensional image is decomposed into

four parts by wavelet transform. Compared with one-dimensional wavelet transform,

two-dimensional wavelet transform decomposes the high-frequency information

component more carefully along the horizontal and vertical directions and further

decomposes the high-frequency part into horizontal detail component, vertical detail

component, and diagonal detail component. Therefore, four parts can be obtained

after the first level decomposition: approximate detail component, horizontal detail

component, vertical detail component, and diagonal detail component. In theory, the

decomposition process can be continued for a long time.

After the multi-scale decomposition of fabric images, wavelet transform can

reconstruct and output approximate and detailed image information according to the

demand. Wavelet has good decomposition and reconstruction ability, so it can obtain

clear and complete fabric structure information and detail information without losing

important information and eliminating interference information.

The principle of automatic detection of fabric warp and weft density is that the

horizontal high-frequency detail component and vertical high-frequency detail

component, which match the height of warp and weft yarn, are obtained by wavelet

decomposition and reconstruction, and the number of warp and weft threads is

calculated to obtain the fabric warp and weft density. The horizontal high-frequency

detail component and vertical high-frequency detail component obtained by different

scales are completely different. The matching degree of detail component and yarn

number is also completely different. Therefore, the decomposition scale of wavelet

decomposition directly affects the horizontal and vertical detail component information

obtained and the accuracy of the fabric warp and weft density information reflected by

the image. To obtain the most complete warp and weft yarn information to obtain the

warp and weft yarn number information matching the image height, the wavelet

decomposition scale has directly impacted the fabric's accuracy. Knowing which scale

is processed is necessary to obtain the reconstructed detail image, which is highly

matched with the fabric image details. We call this wavelet decomposition scale the

optimal decomposition series. Therefore, the research on determining the optimal

decomposition series is of great significance for the automatic calculation and

detection of warp and weft density.

3.3. DETERMINATION OF OPTIMAL WAVELET

DECOMPOSITION SERIES OF FABRIC IMAGE

The correlation coefficient is used to study the degree of linear correlation between

two variables. The curve of the correlation coefficient can reflect the correlation

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

297

between two variables, but it is unclear about the degree of correlation, so it is a kind

of uncertain relationship. To determine the optimal wavelet decomposition series by

the method of correlation coefficient curve is to compare the correlation degree

between the fabric image reconstructed by wavelet decomposition and the fabric

image before decomposition and reconstruction, and calculate the correlation

coefficient of the two images, and take the decomposition series corresponding to the

maximum correlation coefficient as the optimal decomposition series. Then the fabric

is decomposed into high-frequency details and series to get the optimal fabric density.

The concept of the correlation coefficient is that assuming that there are two matrices

of the same dimensions, A and B, the correlation coefficient of a and B is calculated

as follows:

(10)

Where m and N represent the rows and columns of the matrix, respectively.

Figure 2. Sample correlation coefficient curve

From the correlation coefficient curve in the above figure, it can be seen that the

correlation coefficient curve of each sample fabric has a maximum peak value, which

indicates that the reconstructed image obtained by wavelet decomposition and

reconstruction under the corresponding decomposition series is most similar to that

before fabric image processing, Therefore, the maximum peak value of correlation

coefficient curve is selected as the optimal decomposition series of wavelet

decomposition and reconstruction. From the curve in the figure above, the maximum

correlation coefficients of fabrics 1, 2, 3, and 4 are 2, 2, 5, and 3, respectively, which

means that the optimal decomposition order of the four fabrics obtained by the

method of correlation coefficient curve is k = 2, k = 2, k = 5, k = 3, that is, for fabric 1

at decomposition scale 2, for fabric 2 at decomposition scale 2, for fabric 3 at

decomposition scale 5When the decomposition scale of fabric 4 is 3, the horizontal

( )( )

( ) ( )

mn mn

m n

mn mn

m n m n

A A B B

− −

⎛ ⎞⎛ ⎞

− −

⎜ ⎟⎜ ⎟

⎝ ⎠⎝ ⎠

∑∑

∑∑ ∑∑

(a)

(b)

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

298

and vertical high-frequency detail components matching the number of warp and weft

yarns can be obtained by wavelet decomposition and reconstruction respectively.

In the study of determining the optimal decomposition order by the correlation

coefficient, it is found that the results obtained by some fabrics are not very ideal. In

the process of a large number of experimental studies, the optimal decomposition

order determined by the maximum correlation coefficient is still inaccurate. For

example, in the treatment of fabric 3 above, the maximum correlation coefficient is 5.

After processing, the matching degree between the detail component and the warp

and weft yarn of the fabric is not large, so a more stable and accurate method to

determine the optimal decomposition order is studied. Solving the optimal

decomposition series is to obtain the optimal detail components after decomposition

and reconstruction. Through the in-depth study and analysis of all the component

information obtained from many experiments, it is found that there is a certain

relationship between the optimal wavelet decomposition series and the information of

all components obtained after decomposition and reconstruction.

In this paper, the concept of energy curve is introduced to determine the optimal

decomposition series of wavelet decomposition and reconstruction. The energy

values among the approximate, vertical, horizontal, and diagonal components

obtained by wavelet decomposition and reconstruction are calculated using the

energy calculation function energy. The program operation formula is: [a, h, V, D] =

wenergy2 (C, s). A is the approximate low-frequency component of the image after

decomposition and reconstruction, h is the horizontal high-frequency detail component

after decomposition and reconstruction, V is the vertical high-frequency detail

component after decomposition and reconstruction, and D is the diagonal high-

frequency detail component after decomposition and reconstruction. The energy curve

is used to determine the optimal wavelet decomposition series. The wavelet

coefficient square is used for each component after decomposition and reconstruction

by the energy function, and then the energy of each component is obtained by

summation, and then the energy proportion of each component is calculated by

normalization. Finally, the relative gradient change of energy is calculated, and the

energy curve is drawn to observe the change of the energy curve. The lowest peak

value of the energy curve is taken as the optimal decomposition order of the wavelet,

and the fabric image is processed by wavelet transform.

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

299

Figure 3. Wavelet decomposition tree

The two-dimensional signal x (fabric image) is decomposed into four components:

A1, H1, V1, and D1, after the first-order wavelet decomposition. The second-order

decomposition further decomposes A1 into four components, namely A2, H2, V2, and

D2. The third-order decomposition decomposes A2 into four components: A3, H3, V3,

and D3. The subsequent decomposition is carried out according to this law, and the

whole wavelet decomposition process is like branching out. Therefore, the

decomposition process and the resulting components can be represented in a tree

view.

From the above wavelet decomposition tree, we can know that all the components

obtained from the two-dimensional image after K (k is a natural number greater than

0) level decomposition include approximate image component AK, horizontal detail

component H1, H2, HK, vertical detail component V1, V2, VK, and diagonal detail

component D1, D2, Dk. If each component of the output after decomposition and

reconstruction is regarded as a matrix, then the information contained in the

component is stored in the matrix data. Assuming that there are n data in each matrix,

each component is written into a matrix in the form of:

(11)

(12)

(13)

(14)

4. SIMPLE WEAVE FABRIC DENSITY TEST RESULTS

AND ANALYSIS

Simple fabric is the original fabric, which is also called basic fabric. It includes plain

weave, twill weave, and satin weave. All fabrics can not be separated from warp and

[ ]

k 1 2 3k k k kn

A a a a a= …

[ ]

k 1 2 3k k k kn

H h h h h= …

[ ]

k 1 2 3k k k kn

V v v v v= …

[ ]

k 1 2 3k k k kn

D d d d d= …

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

300

weft. This paper's fabric density detection method can monitor all fabrics containing

warp and weft. However, there is a problem of high relative error in the density

monitoring of individual fabrics. To get a better result of fabric design, we need to use

the computer to calculate the fabric density automatically.

4.1. CALCULATION OF WARP AND WEFT DENSITY OF SIMPLE

WEAVE FABRIC PROCESSED BY COMPUTER

From the analysis in the previous section, it can be seen that clear warp and weft

density yarns are arranged. The black horizontal line represents the yarn, and the

white horizontal line represents the gap between the yarn and the yarn. Therefore, the

warp and weft density of the simple woven fabric is required. The number of warp and

weft yarns can be obtained by calculating the number of black transverse lines in the

vertical and horizontal directions. Then the warp and weft yarn density is calculated.

The vertical detail component diagram of fabric represents the arrangement of warp

yarn, alternately arranged between warp and blank space. In the image, the alternate

arrangement of black pixels and white pixels is shown; The horizontal detail

component of the fabric represents the weft arrangement of the fabric.

Similarly, the alternate arrangement of the weft and the blank space means that the

black pixels and white pixels in the vertical direction of the image are arranged

alternately. Therefore, black and white pixels alternate once, representing a yarn. To

calculate the number of yarns in the detail image, we can get it by calculating the

number of consecutive black pixels in the image.

Taking the vertical detail component of fabric as an example, the unit length of

fabric is set as 10cm, the unit of fabric density is the national standard unit "root /

10cm", and the image width of fabric is d (unit is a pixel). Counting the total number of

consecutive black pixels in a row of vertical detail component maps, the yarn number

SJ in the horizontal direction can be obtained. The total number of continuous black

pixels SJ is divided by the width of fabric image D. According to the parameters of the

CCD industrial camera, the camera's resolution can be obtained. Through the

resolution, the number of pixels per centimeter P (in pixels/cm) can be obtained.

Finally, the warp density converted into a standard unit can be obtained by multiplying

MJ and P and multiplying by unit length 10. The calculation formula of fabric warp

density PJ is as follows:

(15)

(16)

If the contour curve is marked as , and all the peak positions of the drop curve

are recorded as and

are the total number of peaks, the weft

density formula of the fabric is expressed as:

(17)

j j

M S d= ÷

j j

P M p=× ×

f(ρ)

ρi,i= 0,⋯,M−1

( )

1 0

2.54

weft

r M

Dρ ρ

−

× −

=× −

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

301

In the formula, is the resolution of the image, and the unit of weft density is/

cm.

After the same preprocessing process, the correlation coefficient curve and energy

curve are used to determine the optimal wavelet decomposition series and the optimal

decomposition series is used to decompose and reconstruct the sample fabric image.

After reconstruction, the vertical and horizontal detail components are optimized, and

the fabric warp and weft density detected by the two methods are calculated.

Table 2. shows the measurement results of the optimal decomposition series determined by

the correlation coefficient curve

No. 1 is a woven fabric, No. 2 is a knitted fabric, and No. 3 is knitted fabric; No. 4 is

non-woven fabric, No. 5 is three-way fabric, No. 6 is multi-directional fabric, No. 7 is

composite fabric, No. 8 is a twill fabric, No. 9 is plain fabric, and No. 10 is checkered

fabric. In manual testing, we use a direct method to measusimple weave fabric's warp

and weft densric. Use fabric density. The number of yarns in 5cm length is measured

by analyzing the mirror, and the warp (weft) yarn density is obtained by multiplying the

number of yarns by 2. Each piece of fabric is measured three times, and the average

value of the three values is taken as the final measurement result of warp and weft

density.

4.2. ANALYSIS OF EXPERIMENTAL RESULTS

For the convenience of measuring the fabric density and weft, we will use the

method of measuring the fabric density and going direct. The accuracy and reliability

of the two methods are analyzed to verify the feasibility of the computer automatic

detection method for the warp and weft density of simple woven fabrics.

Dweft

Fabric number

Warp density (PCS / 10cm)

for the first

time

The second

time

third time average value

1 351.20 355.58 350.12 352.30

2 769.48 771.72 775.19 772.13

3 104.11 107.93 103.14 105.06

4 308.56 311.69 310.56 310.27

5 215.24 210.70 213.09 213.01

6 241.82 236.17 240.12 239.37

7 527.67 535.18 538.30 537.05

8 407.66 401.97 407.68 405.77

9 682.00 681.43 675.13 679.52

10 637.11 640.15 634.85 637.11

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

302

Figure 4. Relative error results

The results show that the relative error of warp and weft density of simple weave

fabric is about 1.00% by using the method of energy curve to determine the optimal

decomposition series. It is found that the relative error of both warp and weft density is

about 1.00%The results show that this method is feasible. However, most of the data

results of the correlation coefficient curve method for determining the optimal

decomposition series are consistent with the results of the energy curve method. The

relative error of the density test results of No. 3 fabric, No. 6 fabric, and No. 7 fabric is

higher than 10%, and the relative error of No. 3 fabric is the highest, reaching 66%.

This shows a serious error in these three fabrics' warp and weft density. That is to say,

the decomposition order determined by the three fabrics' correlation coefficient curves

is not the fabric's optimal decomposition series. Therefore, the fabric warp and weft

density results calculated by the vertical and horizontal detail components obtained by

the wavelet decomposition and reconstruction of the fabric by using this

decomposition series will be so different from the real density of the fabric. This group

of experimental results also further shows that when using wavelet decomposition and

reconstruction methods to detect the fabric warp and weft density, the determination

of the optimal wavelet decomposition series is very important. The results of different

series decomposition and the real warp and weft density of fabric may be greatly

different. Such a large deviation will also greatly impact the subsequent production of

fabrics. Compared with the correlation coefficient curve, the accuracy of the energy

curve method is better. However, in the experiment, it is also found that the

processing error of simple weave fabric with large warp and weft density or complex

pattern and color of the fabric itself may be large, which needs further research and

improvement.

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

303

5. CONCLUSION

This paper aims to use computer vision and digital image processing technology to

replace manual identification and automate simple weave fabric production detection.

This research combines textile knowledge, computer technology, and image

processing technology and achieves the integration of subject knowledge, which is of

great significance for bringing innovative results. To solve the problems existing in the

previous research on the automatic detection of fabric warp and weft density, this

paper puts forward the corresponding algorithm. The main research contents and

innovations are as follows

1.T

he method in this paper combines the correlation coefficient curve method to

determine the optimal decomposition order of fabrics, which are respectively k=2,

k=2, k=5, k=3. When the decomposition scale of fabric 1 is 2, the decomposition

scale of fabric 2 is 2, the decomposition scale of fabric 3 is 5, and the

decomposition scale of fabric 4 is 3. Therefore, this method can effectively avoid

warp and weft skew.

2. Only the fabric density detection method in this paper is used. The relative error of

the warp and weft density of the simple woven fabric detected is 1%, which is

consistent with the curve with the most decomposition sequence, indicating that the

relative error detected by the fabric density detection method in this paper is the

smallest and the accuracy is the highest.

3. In this paper, the fabric density detection method decomposes and reconstructs the

horizontal, vertical, and high-frequency detail components and approximate detail

components of the fabric image. Most of the data from the optimal decomposition

sequence determined by the correlation coefficient curve method are consistent

with the results obtained by the energy curve method. However, there is a problem

of high relative error in the density monitoring of individual fabrics.

DATA AVAILABILITY

The data used to support the findings of this study are available from the

corresponding author upon request.

CONFLICT OF INTEREST

The authors declare that the research was conducted without any commercial or

financial relationships that could be construed as a potential conflict of interest.

REFERENCES

(1) Shukla, K., Ahmad, A., Ahluwalia, B. S., et al. (2022). Finite element simulation

of transmission and reflection of acoustic waves in the ultrasonic

transducer. Japanese Journal of Applied Physics.

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

304

(2) Elemmi, M. C., Anami, B. S., & Malvade, N. N. (2021). Defective and

nondefective classification of fabric images using shallow and deep

networks. International Journal of Intelligent Systems.

(3) Raj, A., Sundaram, M., & Jaya, T. (2020). Thermography based breast cancer

detection using self-adaptive gray level histogram equalization color

enhancement method. International Journal of Imaging Systems and

Technology.

(4) Ojo, J. F., & Olanrewaju, R. O. (2021). Review of Family of Autoregressive

Integrated Moving Average Models in the Comportment of Autocorrelation

Function for Non-Seasonal Time Series Data. International Journal of

Mathematical Sciences & Applications, 19(1), 79-89.

(5) Wu, D., & Tang, Y. (2020). An improved failure mode and effects analysis

method based on uncertainty measure in the evidence theory. Quality and

Reliability Engineering, (1).

(6) Cao, Y., You, J., Shi, Y., et al. (2021). Research on the Green

Competitiveness Index of Manufacturing Industry in Yangtze River Delta

Urban Agglomeration. Problemy Ekorozwoju, 16(1), 143-156.

(7) Trafton, K., & Giachetti, T. (2021). The morphology and texture of Plinian

pyroclasts reflect their lateral sourcing in the conduit. Earth and Planetary

Science Letters, 562.

(8) Le, B., Troendle, D., & Jang, B. (2021). Detecting fabric density and weft

distortion in woven fabrics using the discrete fourier transform.

(9) Kaplan, V. (2021). Detection of Remote Sensing Warp Tension during

Weaving on Plain Twill and Satin Fabric. Fibres and Textiles in Eastern

Europe, 29(1(145)), 35-39.

(10) Monfared, S. S., & Sedef, B. (2021). Road Lane detection through image and

video processing using edge detection and Hough transform for

autonomous driving purposes.

(11) Xz, A., Gla, B., Zya, B., et al. (2020). Parameter estimation based on Hough

transform for airborne radar with conformal array - ScienceDirect. Digital

Signal Processing, 107.

(12) Qin, T., Cao, P., Zhang, Y., et al. (2021). Underwater magnetic target signal

denoising based on modified wavelet decomposition and reconstruction

algorithm. Journal of Physics: Conference Series, 1738(1), 012019 (8pp).

(13) Yan Y, Liu Y, Yang M, et al. (2020). Generic wavelet-based image

decomposition and reconstruction framework for multi-modal data

analysis in smart camera applications. IET Computer Vision, 14(7), 471-479.

(14) Lgt, A., Mld, B., & Jgg, C. (2022). Picking out the warp and weft of the

Ediacaran seafloor: Paleoenvironment and paleoecology of an Ediacara

textured organic surface. Precambrian Research, 369, 106539-.

(15) Wang, X., Wang, S., Guo, Y., et al. (2020). Research on improved sharpening

algorithm based on closed operation and binarization. Journal of Physics

Conference Series, 1629, 012019.

(16) Shi, J., Wang, Y., Zhang, X., et al. (2021). Extraction method of weak

underwater acoustic signal based on the combination of wavelet transform

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

305

and empirical mode decomposition. International Journal of Metrology and

Quality Engineering.

(17) Nugroho, P.C., Widadi, R., & Zulherman, D. (2021). Hand and Foot Movement

of Motor Imagery Classification Using Wavelet Packet Decomposition and

Multilayer Perceptron Backpropagation. In 2nd Borobudur International

Symposium on Science and Technology (BIS-STE 2020).

(18) Pavicˇic, I., Brievac, Z., et al. (2021). Geometric and fractal characterization

of pore systems in the upper Triassic dolomites based on image

processing techniques example from umberak mts nw Croatia.

Sustainability, 13.

(19) Pang, K., Alam, M. Z., Zhou, Y., et al. (2021). Adiabatic Frequency Conversion

Using a Time-Varying Epsilon-Near-Zero Metasurface. Nano Letters.

(20) Tang, S.C., Wang, X.Y., Yang, L.L., et al. (2020). Design of Slotted Dielectric

Patch Antenna with Filtering Characteristic. In 2019 International Symposium

on Antennas and Propagation (ISAP). IEEE.

(21) Tutatchikov, V. (2020). Application of parallel version two-dimensional fast

Fourier transform calculating algorithm with an analogue of the Cooley-

Tukey algorithm. In 2020 International Conference on Information Technology

and Nanotechnology (ITNT).

(22) Niu, H., Wu, B., Wang, Q., et al. (2020). Research on steel barrel flattened

seam recognition based on machine vision. Journal of Physics Conference

Series, 1633, 012014.

(23) Barreto, M., Reis, J., Muraoka, T., et al. (2021). Diffuse reflectance infrared

Fourier transform spectroscopy for a qualitative evaluation of plant leaf

pigment extraction.

(24) Shi J, Wang Y, Zhang X, et al. (2021). Extraction method of weak underwater

acoustic signal based on the combination of wavelet transform and

empirical mode decomposition. International Journal of Metrology and Quality

Engineering.

(25) Le B, Troendle D, Jang B. (2021).

Detecting fabric density and weft distortion

in woven fabrics using the discrete Fourier transform.

(26) Tabunschik V. A., Т. M. Chekmareva, & Gorbunov R. V. (2020). Spectral

characteristics of some agricultural crops in different phenological phases

of vegetation. Plant Biology and Horticulture Theory Innovation, 152, 56-70.

(27) Cooney G. S., Barberio M., Diana M., et al. (2020). Comparison of spectral

characteristics in human and pig biliary system with hyperspectral

imaging (HSI).

(28) Bo C., Polatkan G., Sapiro G., et al. (2020). The hierarchical beta process for

convolutional factor analysis and deep learning. In ICML.

(29) Kaseng, F., Lezama, P., Inquilla, R., & Rodriguez, C. (2020). Evolution and

advance usage of Internet in Peru.

3C TIC. Cuadernos de desarrollo aplicados

a las TIC, 9(4), 113-127. https://doi.org/10.17993/3ctic.2020.94.113-127

(30) Liu Chunguang. (2021). Precision algorithms in second-order fractional

differential equations. Applied Mathematics and Nonlinear Sciences, 7(1),

155-164. https://doi.org/10.2478/AMNS.2021.2.00157

https://doi.org/10.17993/3ctic.2023.121.287-306

3C TIC. Cuadernos de desarrollo aplicados a las TIC. ISSN: 2254-6529

Ed.42 | Iss.12 | N.1 January - March 2023

306