A computer vision system for objective fabric smoothness appearance assessment with an ensemble classifier

Abstract

Fabric smoothness appearance assessment plays an important role in the textile and apparel industry. To evaluate fabric smoothness objectively, different methods have been proposed based on computer vision technology. To further improve the performance and promote the application of the assessment methods, this paper reports a hybrid computer vision system for objective assessment of fabric smoothness appearance with an ensemble classifier to integrate the advantages of the different image feature sets, which are extracted based on different image processing technologies. The image acquisition environment is established in this system with the selection of illumination parameters—intensity, position angle and altitudinal angle—by a designed strategy. The main steps of the strategy include determination of priority by information gain analysis and parameter selection by classifier performance analysis. The support vector machine classifiers trained by each feature sets are grouped into an ensemble by a self-adapting weighted voting method and the redundant feature sets are eliminated based on the weights of the feature sets. The final result shows evaluation accuracies with 82.86% under 0-degree error, 97.14% under 0.5-degree error and 100% under 1-degree error, which outperforms the other methods in the same environment and verifies the applicability of the proposed system.

Keywords

Fabric smoothness textile testing ensemble classifier computer vision system

Fabric smoothness after laundering is regarded in the textile and garment industry¹ as a vital characteristic of a fabric when evaluating the tendency of the fabric to wrinkle. The quantity of wrinkles on a fabric after being subjected to laundering procedures has a bearing on its ‘ease-of-care’ related properties (durable press, easy care, minimum iron, after-wash appearance, etc.). Conventionally, the smoothness of the fabric samples after standard laundering is assessed visually in accordance with smoothness appearance replicas in six degrees of smoothness (SA-1, SA-2, SA-3, SA-3.5, SA-4, SA-5) referring to the relevant standards.^2,3 However, human vision is affected by individual physical, psychological and environmental factors, with low precision and poor reproducibility.⁴ To avoid such limitations, objective methods of evaluation have been proposed by researchers in recent years. Previous research has treated the assessment of fabric smoothness as an object classification problem, which generally includes three main steps: acquisition of data on specimen surface, extraction of appearance features and classification of degree of smoothness.

According to the data forms, two types of methods exist in this field: two dimensional (2D) and three dimensional (3D). The 2D methods are based on 2D digital images of the fabric samples captured by a scanner or an industrial camera. The imaging environment of scanners is stable, but their vertical incident light weakens the gray-scale changes that represent the fabric wrinkles. Integrated with the non-vertical incident light, industrial array cameras can capture images with rich wrinkle features under conditions similar to the manual method. As the 2D image of the fabric is a projection of its depth map from the world space to the imaging space going through the effect of light and shadow and the sensor imaging, the 2D method is a kind of indirect observation. As a type of direct observation, on the other hand, 3D methods utilize 3D depth maps of the fabric samples captured by the technologies of laser triangulation,^5,6 photometric stereo vision^7,8 and binocular stereo vision^9,10 as the raw data. These 3D methods have recently attracted more attention in the field, however, they suffer a number of inevitable drawbacks such as poor efficiency, high cost and high calibration complexity. In the authors' opinion, although 2D methods can only provide indirect observations of the fabric surface, they can still provide sufficient information for evaluation of fabric smoothness, due to the phenomenon that a trained tester can well assess a specimen with only 2D images observed by eye. The main problem that needs to be solved initially is to determine a highly adaptable image acquisition environment which provides the most reliable foundation for the further stable and effective feature extraction and smoothness classification.

Based on the captured data, a number of features based on different technologies are applied to describe the smoothness of fabrics. The features are listed below, based on the different data forms:

Features from 2D images: edge area; shade area¹¹; variation of the gray-level intensity¹²; angular second moment, contrast, correlation, entropy of the gray-level co-occurrence matrix^13,14; gray-scale range around the edges^15,16; the fast Fourier transform subtotal in a specific frequency range¹⁷; orientation, hardness, density and contrast of wavelet coefficients.¹⁸

Features from 3D depth maps: length, surface area, volume under the surface, mean principal curvatures, mean maximum twist of every sub-block of the depth map⁵; arithmetic average roughness, root mean square roughness, 10-point height, bearing surface ratio, wrinkle sharpness, wrinkle density^6,19; fractal dimension²⁰; maximum amplitude, sharpness, density, maximum amplitude of the first derivative of the cross profile of the edges^10,21; mean, mean deviation, and standard deviation of the height values in each row.²²

For the classification models, minimum distance algorithm,¹⁷ fuzzy priority similarity comparison method,¹⁴ neural network,^6,13 support vector machine (SVM),^4,18 are widely used. In addition, differing from the traditional machine learning algorithms which depend heavily on the representation of the input image, deep learning algorithms can extract the high-level, abstract features from the input images by introducing representations that are expressed in terms of other simpler representations.²³ Such algorithms have been widely used in the field of image classification, for example, of ships²⁴ or vehicles,²⁵ hyperspectral image classification,²⁶ and candidate classification in detection of lung nodules.²⁷ However, the performance of deep learning algorithms depends heavily on the amount of training data, which is not well satisfied due to the small sample in this research. In the experiments conducted for this research, the performances of a series of well-known deep learning models are discussed.

As introduced above, the authors infer that 2D methods are sufficient to solve the problem. To explore further the possibility of the application of 2D methods in fabric smoothness assessment, the image acquisition environment needs to be further considered and it needs to be controlled because the environment greatly influences the visibility of wrinkle features in images. In addition, although several series of image features have been proposed independently in this area, the combination and comparison of different image features, and the possibility of improvement of the classifier ensemble, can be further explored. This research proposes a 2D-image-based system for the objective assessment of fabric smoothness with a synthesis discussion on the determination of the image acquisition environment, the selection of features and the classifier ensemble. First, an image acquisition system was established with a mechanism designed to capture images in different illumination environments. A number of feature sets based on different image processing technologies were then selected and used to train the SVM models. To determine a highly adaptable illumination environment, the information gain was used to select an initial environment and assign priorities to the parameters, that is, the position angle, altitudinal angle and intensity of the illumination, which were then selected by consideration of the classification accuracy and environment stability. Finally, a self-adapting weighted voting method was proposed to eliminate redundant feature sets and construct a classifier ensemble which can obtain a better result than any single feature sets or classifiers.

The contribution of this research is as follows:

Considering the inevitable disadvantages of the 3D methods, the feasibility of the more efficient and economical 2D methods is discussed and verified, promoting the further industrial application of the objective fabric smoothness assessment system.

The image acquisition environment is widely discussed for different image features to evaluate the fabric smoothness, which provides a new reference for further research in this area.

Based on the previous studies, a series of image feature sets is concluded and proposed. An ensemble classifier is established, which selects and combines the different feature sets and was found to outperform every individual classifier trained by every single feature set.

Method

In this section, the constitution of the proposed system is introduced, which includes the image acquisition system, image preprocessing, features extraction, classification model, feature evaluation, feature selection and classifier ensemble.

Image acquisition system

The main purpose of this research is to propose a 2D-image-based fabric smoothness assessment system with high stability across different conditions, for which the basic research data set is an image set of fabric specimens under different illumination environments. An image acquisition system able to capture images of the specimens under numerous different light conditions was designed. As shown in Figure 1, the whole system is set up in a light box. An objective table is set up at the bottom center of the light box, with a CCD camera (Point Grey Chameleon CMLN-13S2C) installed vertically above it to capture images of the specimen. To control the light environment, a light control module, including a strip-LED light source, two step-motors, a rotating arm and a screw rod, is built around the center pole of the objective table. As shown in Figure 1, step-motor A controls the light source on the rotating arm to rotate about the specimen horizontally, and step-motor B controls the light source to move up and down vertically. With cooperation between step-motor A and step-motor B, the light source can be moved in space. Additionally, with the help of the light source controller, images of the specimen can be captured under different light conditions with the combination of different position angle, altitudinal angle and intensity of the light source. To avoid interference from reflected light inside the light box, the internal surface of the light box was covered with black flocking fabric. Using this image acquisition system, images of a set of fabric samples were captured under different illumination environments, as illustrated in Figure 2.

Figure 1.

Image acquisition system for 2D images of fabric samples.

Figure 2.

Image of a fabric sample captured by the proposed image acquisition system under different environments: column (1): illumination intensity change; column (2): position angle change; column (3): altitudinal angle change.

Image preprocessing

To rectify the uneven illumination background in images caused by the one-sided light source, an image preprocessing operation is applied to the original image I as follows:

F = I / I_{f}

(1)

where

I_{f}

is the 2D binomial fitting of I. Using this process, an uneven illumination background can be rectified well.

Features extraction

To study the performance of different kinds of features extracted from the fabric images, eight different feature sets were selected by reference to the literature and developed by the authors.

1. Edge-based features

First, the Canny edge operator is applied to an input image I to generate an edge binary image $BW$ . Then the gray-level ranges of the pixels in a 5 × 5 pixel window centered at the edge pixels are calculated and grouped into a vector $v$ .

Edge-based features set 1

As proposed in previous research,¹⁶ three statistical features: the mode of $v$ , the mean value and the standard deviation of the distribution histogram of $v$ are calculated and grouped into the first edge-based features set, namely $F_{edge}^{1}$ .

Edge-based features sets 2 and 3

In this study, to discuss further the information in $v$ , the histogram of $v$ and a set of percentiles are tested as two independent feature sets, that is, $F_{edge}^{2}$ and $F_{edge}^{3}$ , which can be formulated as follows:

F_{edge}^{2} (n) = \sum_{i} f_{c} (v (i)) n = 1, 2, 3, \dots, L / t

(2)

where

f_{c} (\cdot)

is a function as

f_{c} (x) = \begin{matrix} \begin{matrix} 1 (n - 1) \cdot t_{1} \leq x < n \cdot t_{1} \\ 0 else \end{matrix} \end{matrix}

(3)

where

t_{1}

is the step length of the histogram, which is selected as 15 in this study;

n = 1, 2, 3, \dots, L / t

is the index of the histogram;

L = 255

is the maximum gray-scale value in an 8bit image; and

F_{edge}^{3} (n) = pct (n \cdot t_{2}) n = 1, 2, 3, \dots, 9

(4)

where

t_{2}

is the step length of the percentile values, which is selected as 10 in this study;

pct (\cdot)

is the percentile calculation function.

Gray-level distribution-based features In another study,¹⁴ a set of features are extracted from the image based on the first-order statistical analysis of the pixel gra- level distribution. The feature set has four elements, as follows: Variance $σ^{2}$

σ^{2} = \sum_{l = 0}^{L} (l - \bar{l})^{2} prob (l)

(5)

Energy E

E = \sum_{l = 0}^{L} prob (l) 2

(6)

Entropy H

H = - \sum_{l = 0}^{L} prob (l) \cdot log 2 prob (l)

(7)

Skewness S

S = (1 / σ^{3}) \sum_{l = 0}^{L} (l - \bar{l})^{3} prob (l)

(8)

where l is the pixel gray level in

{0, 1, 2, \dots, L}

; L is the maximum gray level;

prob (l)

is the probability of the gray level l in the image. Thus, the feature set based on the gray-level distribution, namely

F_{gray}

, can be extracted as follows:

F_{gray} = [σ^{2}, E, H, S]

(9)

Gray-level commensal matrix-based features Based on the gray-level commensal matrix (GLCM) on direction θ of the fabric image, namely $K^{θ}$ , four features are calculated for texture analysis: consistency $Q_{1}^{θ}$ , contrast $Q_{2}^{θ}$ , entropy $Q_{3}^{θ}$ , and gray occurrence $Q_{4}^{θ}$ .^13,14 In this paper, the GLCM direction $θ_{i}$ is valued as the vector $[0^{\circ}, 45^{\circ}, 90^{\circ}, 135^{\circ}]$ respectively. Thus the GLCM-based feature set $F_{glcm}$ is generated as follows:

F_{glcm} (n) = Q_{i}^{j} n = 1, 2, 3, \dots, 16

(10)

where

n

is the index of the element in

F_{glcm}

;

i = 1, 2, 3, 4

and

j = 1, 2, 3, 4

is the index of the features introduced above.

Fourier spectrum-based features Using the fast Fourier transform (FFT), the input image can be converted from the spatial domain into the discrete Fourier frequency domain.^17,28 Fourier spectrum amplitude in some specific frequency ranges was calculated as the feature to evaluate the fabric smoothness in a previous study.¹⁷ In this paper, a set of ranges covering the whole Fourier spectrum is considered, in which the amplitudes are calculated and grouped as the feature set $F_{fft}$ , which can be formulated as follows:

F_{fft} (n) = \sum_{u = 0}^{M - 1} \sum_{v = 0}^{M - 1} F^{2} (u, v) \cdot h_{n} (u, v) n = 1, 2, 3, \dots, N

(11)

where

F (u, v)

is the Fourier spectrum of the fabric image generated by FFT;

h_{n}

is a set of band-pass filter defined as

\begin{matrix} h_{n} (u, v) = \begin{matrix} \begin{matrix} 0 D_{n} - \frac{w}{2} \leq D (u, v) < D_{n} + w / 2 \\ 1 \end{matrix} n = 1, 2, 3, \dots, N \end{matrix} \end{matrix}

(12)

where w is the band-width of the filter which is selected as 10 in this study;N is the total number of the band-width filters which is selected as 19 in this study; D_n is the center diameter of the nth band-pass filter valued as

D_{n} = n \cdot w

Wavelet-based features: Wavelet transform (WT) can decompose the image into different frequency domains on different directions and compose their wavelet coefficients into a pyramid-structured wavelet decomposition.¹⁸ To discuss the performance of the wavelet-based features, in this paper, two feature sets are selected to be used in the experiment.

Wavelet-based-features set 1

The fabric image is processed by a 2D Haar WT operation and a pyramid-structured decomposition is generated, in which every rectangle $C^{l, d}$ illustrates the coefficients in a specific frequency level l and on a specific direction d. As in the study,¹⁸ the wavelet-based features set 1 $F_{wavelet}^{1}$ consists of the total energy ( $En$ ), wrinkle hardness ( $Hd$ ), wrinkle density ( $De$ ) and the contrast parameter ( $Cn$ ) in levels 4 and 5 in all the directions.

Wavelet-based-features set 2

In addition to the $F_{wavelet}^{1}$ , to discuss the full possibility of the features in WT coefficients, a wider-domain-considered features set $F_{wavelet}^{2}$ is generated as the mean value of the coefficients $C^{l, d}$ in every level l and on every direction d.

F_{wavelet}^{2} = [mean ((C^{l, d}) 2) | l = 1, 2, \dots 6; d = 1, 2, 3]

(13)

Classification model

This paper has shown how the feature sets can be extracted from an input fabric image. To further acquire the smoothness degree of the specimen and evaluate the performance of the different feature sets, the feature sets of the images in the dataset are trained by the support vector machine (SVM) model. For a training set of instance-label pairs $(x_{i}, y_{i}), i = 1, 2, \dots, l$ where $x_{i} \in R^{m}$ and $y_{i} \in {1, - 1} l$ , the objective function of the SVM model is defined as follows²⁹:

\begin{matrix} {min}_{w, b, ξ} \frac{1}{2} w^{T} w + C \sum_{i = 1}^{l} ξ_{i} \\ subject to y_{i} (w^{T} φ (x_{i}) + b) \geq 1 - ξ_{i}, \\ ξ_{i} \geq 0 \end{matrix}

(14)

where training vectors

x_{i}

are mapped into a higher dimensional space by the function

φ (\cdot)

;

C > 0

is the penalty parameter of the error term; and

K (x_{i}, x_{j}) \equiv φ (x_{i})^{T} φ (x_{j})

is named as the kernel function. In this paper, the radial basis function is used as the kernel function, which is formulated as follows:

K (x_{i}, x_{j}) = exp (- γ ‖ x_{i} - x_{j} ‖ 2), γ > 0

(15)

where γ is the kernel parameter. In this paper, the model parameters C and γ are optimized by the grid search method based on five-fold cross-validation accuracy.

Verifiable indicator

As proposed above, different kinds of features can be extracted from the fabric images. To evaluate the effectiveness and stability of the features, a specific index has to be used. In this paper, information gain, which is widely used in the feature selection field,³⁰ is used as such an index. The feature data set is described as D, in which every row describes a sample and every column describes a specific feature. When the i^th feature a is being discussed, the value of a can divide D into V subsets, thus the information gain $G (a)$ can be formulated as follows:

G (a) = Ent (D) - \sum_{v = 1}^{V} \frac{| D^{v} |}{| D |} Ent (D^{v})

(16)

Ent (D) = - \sum_{l = 0}^{L} prob (l | D) \cdot log 2 prob (l | D)

(17)

where l is sample label in

{1, 2, \dots, L}

; L is the maximum sample label; and

prob (l | D)

is the proportion of the samples that belong to the class l in D.

Feature selection and classifier ensemble method

Based on the above feature sets and the SVM classifier, several different classifiers can be generated. Finally, to generate the best performing classification result, in this paper, a weighted voting classifier ensemble method is proposed based on linear regression, which includes two main steps: weight determination and elimination of feature sets.

Weight determination

Note that n SVM classifiers have been trained based on the above n feature sets, which generate predictions for the m samples comprising an $m \times n$ matrix P with every element $P_{i, j}$ corresponding to the prediction of the i^th sample and j^th classifier. Thus the target of the ensemble method is to determine a weight vector $w \in R^{n \times 1}$ for every classifier which maximizes the similarity between the weighted voting result and the label of the samples $L \in R^{m}$ . Due to the fact that the voting result involving the calculation of mode is hard to formulate, and to the phenomenon that all the class labels (fabric smoothness levels) in this problem are value-meaningful (e.g. the smaller the smoothness level prediction error, the better the classifier performs), the objective function can be replaced by a relaxation function as follows:

min ‖ P w - L ‖ 2

(18)

This objective function is equivalent to the linear regression problem whose closed-form solution can be obtained.

Elimination of feature sets

After the weight vector has been determined, the weighted voting can be implied. However, the performance of the classifiers ensemble can be further discussed by eliminating some redundant feature sets. For this purpose, a greedy algorithm can be used, which eliminates the least-weighted feature set one by one and evaluates performance of the remaining ensemble. Finally, the classifiers ensemble can be selected based on the consideration of feature simplification and accuracy of final classification.

Experiment

Experiment setup

With the exception of the image acquisition system describe above, all the experiments in this study were conducted on a personal computer with Intel(R) Core(TM) I7-4790 CPU(3.6 GHz) 16GB RAM and GPU of Nvidia(R) GTX 1080Ti. All the algorithms were implemented by MATLAB under the Windows 10 and Linux operating system. The SVM was implemented by the LIBSVM.²⁹ The deep learning algorithms were implemented by the PyTorch.³¹

Materials

To evaluate the performance of the proposed fabric smoothness assessment system, four different fabrics of different weave structure and fiber composition were selected from a textile mill and cut into pieces of 380 mm × 380 mm according to the AATCC standard.² There were 385 samples in total. The resized samples were then laundered in different modes, with 2003 AATCC Standard Reference Liquid Detergent and 25^oC rinse temperature, to generate diverse fabric smoothness appearances. The samples were then graded by experts. Finally, images of the samples were captured by the proposed image acquisition system in different illumination environments. Detailed information on the final fabric data set is given in Table 1. As introduced above, the proposed image acquisition system is able to acquire images with different illumination intensities, position angles and altitudinal angles. In this research, such parameters were set as a series of discrete value, as given in Table 2, and images of each fabric sample were collected under different combinations of the three parameters. For the sake of simplicity, in further experiments, the value of the parameters is described as their sequence number in Table 2, and the combination of the altitudinal angle and the illumination intensity is expressed as the joint of their sequence number such as environment 1-3, which means the image acquisition environment is set as altitudinal angle 1 and illumination intensity 3.

Table 1.

Information on fabric samples and number of samples in different degrees of smoothness appearance

Fabric				Sample size
Weave	Fiber content	Yarn count	Weave density (thousands of in.)	SA-1	SA-1.5	SA-2	SA-2.5	SA-3	SA-3.5	SA-4	SA-5
Plain	100% C	40S×40S	117 × 79	25	13	15	15	7	7	10	6
Plain	65% C 35% P	32S×32S	75 × 60	2	1	3	4	7	25	25	29
Twill	100% C	50S×50S	139 × 90	18	12	13	15	9	11	12	4
Twill	65% C 35% P	32S×32S	77 × 58	3	2	1	4	6	24	23	34

Table 2.

Parameters in the image acquisition system

Sequence number	Illumination intensity	Position angle	Altitudinal angle
1	100	0°	27°
2	150	90°	45°
3	200	180°	56°
4	250	270°	63°
5	–	–	68°

In the experiments, five-fold cross-validation was applied to evaluate the classification performance reasonably. In this process, the whole data set was uniformly divided into five subsets, each of which was treated as the testing set in turn with the others being treated as the training set. The final result was computed by the average result across the five subsets. Thus, the training sample size was 308 and the testing sample size was 77 on average for each subset.

Results and discussion

In this research, the main purpose is to establish a whole system including determination of image acquisition environment, selection of feature sets and classifier ensemble, which will finally result in a stable and efficient solution for the objective assessment of fabric smoothness.

Determination of the image acquisition environment includes three main parameters of illumination: intensity, position angle, and altitudinal angle. In the experimental process, the illumination intensity and altitudinal angle was set in a specific value and the illumination position angle was selected as the number of illumination directions, that is, determining the number of fabric images with different illumination position angles to be used to extract features.

Initial environment selection

In the first experiment, the performance of the different feature sets in different image acquisition environments was first evaluated by the mean information gain (MIG) of every element feature in the set. Based on the value distribution of the MIG, an initial environment for the further determination can be determined roughly by selecting an environment with a relatively high MIG, for example, altitudinal angle 3 and illumination intensity 2, or environment 3-2. Although the value of the MIG does not describe the exact final performance of the feature set, it helps to evaluate the stability of the features in different illumination conditions. As given in Table 3, the variances of MIG (VMIG) at different altitudinal angles of illumination were much larger than for different illumination intensities, which means the performance stability of all the feature sets is higher with changes in the illumination intensity. Thus, in the further experiments, the selection of altitudinal angle of illumination was set prior to the illumination intensity.

Table 3.

VMIG of feature sets under different conditions

Feature sets	VMIG
Feature sets	Altitudinal angle	Illumination intensity
$F_{edge}^{1}$	32.7526	0.3007
$F_{edge}^{2}$	0.3427	0.0349
$F_{edge}^{3}$	2.663	0.0403
$F_{fft}$	0.1173	0.0059
$F_{wavelet}^{1}$	1.8078	0.0353
$F_{wavelet}^{2}$	1.0055	0.1013
$F_{gray}$	1.8925	0.0986
$F_{glcm}$	0.3715	0.0068

Determination of parameters of the image acquisition system

Because wrinkles appear randomly on the fabric surface, much information can be neglected by feature sets which only consider the features of a fabric image with a single illumination position angle. Thus, the experiment for illumination direction number was first implemented in the initial environment, in which the classification accuracies of the SVM trained on different feature sets with features extracted from fabric images with

n_{p} = 1, 2, 3, 4

illumination directions were calculated. As shown in Table 4, the classification accuracy increases significantly when the number of illumination directions is increased from one to two, but shows little increment after that. Considering the equipment cost and the feature vector dimension, two is the optimal choice of the number of illumination directions selected to capture images to extract feature sets.

Table 4.

Performance of the feature sets with different numbers of illumination direction angles

Feature sets	Classification accuracy (%)
Feature sets	One side	Two sides	Three sides	Four sides
$F_{edge}^{1}$	64.68	77.92	78.18	78.96
$F_{edge}^{2}$	63.12	73.77	74.03	75.58
$F_{edge}^{3}$	65.45	78.7	77.14	77.92
$F_{fft}$	68.83	77.66	77.14	77.14
$F_{wavelet}^{1}$	64.68	75.58	74.03	72.99
$F_{wavelet}^{2}$	71.95	77.4	77.92	78.7
$F_{gray}$	64.94	68.83	70.91	71.95
$F_{glcm}$	70.13	73.51	72.47	72.21

The performance of the classifiers trained on different feature sets in a specific environment can be used to verify the effectiveness of the model in such an environment, however, it is not enough to demonstrate the environmental stability and robustness. Thus, in the experiments, SVM classifiers trained on the feature sets extracted from one specific environment were tested on both the specific environment and the corresponding neighboring environments. For instance, the neighboring environments of environment 3-2 are environment 3-1 and 3-3 when discussing the illumination intensity; and they are environment 4-2 and 4-2 when discussing the altitudinal angle. In addition, to evaluate the maximum assessment possibility of the feature sets in a specific environment, the extreme maximum classification accuracy $a_{m}$ was used, namely environment accuracy, which can be calculated as the ratio of the number of samples that are correctly classified by any single classifier to the total number of the samples.

In a further experiment, based on the priority determined in the previous experiment, the altitudinal angle was discussed first. As shown in Table 5, the classification accuracy (Acc) and the maximum classification accuracy reduction in the neighboring environments (dAcc) were calculated. The Acc of the feature sets shows relatively high diversity as the altitudinal angle of illumination changes and rather low environmental stability. Generally, as the altitudinal angle increased, the Acc and dAcc decreased. Synthetically considering the environment accuracy for all the feature sets, environment 4-2 performed a relatively high environment Acc and a lowest environment dAcc which means the best environment stability. Hence it was selected as the environment for the selection of illumination intensity.

Table 5.

Classification accuracy and environment stability performance of the feature sets according to change of altitudinal angle of illumination

Feature sets	Classification Acc/dAcc (%)
Feature sets	1-2		2-2		3-2		4-2		5-2
$F_{edge}^{1}$	76.36	20.26	76.88	19.48	78.96	15.06	76.88	13.50	77.92	23.11
$F_{edge}^{2}$	77.66	25.45	76.36	20.00	77.4	10.13	75.84	7.79	73.77	0.52
$F_{edge}^{3}$	79.48	49.35	79.22	58.96	80	35.84	78.44	21.82	76.88	18.44
$F_{fft}$	80.26	25.2	79.22	16.10	78.96	6.75	78.96	3.90	78.44	4.93
$F_{wavelet}^{1}$	71.95	11.95	72.21	10.91	70.91	12.21	70.65	9.09	70.91	3.64
$F_{wavelet}^{2}$	78.18	27.53	76.1	15.58	77.14	4.41	77.14	5.19	77.14	3.89
$F_{gray}$	72.21	8.31	75.06	18.96	75.32	23.89	73.77	14.03	74.55	9.10
$F_{glcm}$	80.26	25.2	78.18	13.50	77.66	24.93	78.7	25.45	79.48	33.51
Environment	96.1	8.05	96.62	7.53	95.58	3.63	95.58	–0.78	94.55	0.00

Based on the selected altitudinal angle, the performance of the classifiers in different illumination intensities was discussed. As shown in Table 6, the diversities of the Acc and dAcc were much lower than those in the altitudinal angle discussion. This tendency is consistent with the prediction in the previous discussion on information gain. Considering the higher classification accuracy and environment stability, environment 4-3 was selected as the ideal environment for its highest environment Acc and relatively low dAcc.

Table 6.

Classification accuracy and environment stability performance of the feature sets according to change of illumination intensity

Feature sets	Environment Acc/dAcc (%)
Feature sets	4-1		4-2		4-3		4-4
$F_{edge}^{1}$	77.92	10.65	76.88	7.27	76.62	0.78	76.36	3.38
$F_{edge}^{2}$	74.29	1.82	75.84	5.19	74.81	2.34	75.84	1.82
$F_{edge}^{3}$	76.62	1.04	78.44	3.38	78.18	2.6	77.92	1.82
$F_{fft}$	79.48	1.04	78.96	1.3	79.22	1.3	78.96	0.52
$F_{wavelet}^{1}$	70.65	2.6	70.65	2.6	72.21	1.82	73.25	3.64
$F_{wavelet}^{2}$	76.62	0	77.14	0	78.44	2.08	77.14	0.78
$F_{gray}$	73.51	3.9	73.77	2.6	74.03	2.08	73.77	0.78
$F_{glcm}$	78.7	1.56	78.7	2.6	77.4	0.26	77.66	1.04
Environment	93.77	0	95.58	1.82	95.58	0.52	94.55	0

Feature selection and classifiers ensemble

When the image acquisition environment had been determined, the ensemble method proposed in the previous section was applied. The weights of the feature sets are given in Table 7, and the performance of the classifier ensemble in the feature sets elimination process is shown in Figure 3. In this process, as the data set was reduced by the proposed method, the classification accuracy of the classifier ensemble increased at first, and then decreased when there were less than five remaining feature sets, which verifies the redundancy of the first three feature sets. Thus the final reserved feature sets were determined, as shown in Table 7. The final classification accuracy of the classifier ensemble reached 82.86% and was much higher than any single classifier in the same environment.

Figure 3.

Classification accuracy of the classifier ensemble as feature sets are eliminated.

Table 7.

Weights and the elimination condition of the feature sets generated by the proposed method

Feature sets	Weight	Elimination condition
$F_{edge}^{1}$	0.1816	Reserved
$F_{edge}^{2}$	0.0394	Eliminated
$F_{edge}^{3}$	-0.0195	Eliminated
$F_{fft}$	0.2537	Reserved
$F_{wavelet}^{1}$	0.1099	Reserved
$F_{wavelet}^{2}$	0.2396	Reserved
$F_{gray}$	0.1555	Reserved
$F_{glcm}$	0.028	Eliminated

Comparison of results of different methods

To verify the effectiveness of the proposed system, based on the same dataset, different previous 2D methods for the fabric smoothness assessment were applied for comparison. Additionally, a group of general deep learning image classification methods were applied in the comparison experiment: LeNet,³² AlexNet,³³ GoogLeNet: Inception V3³⁴ and ResNet50.³⁵ The hyperparameters were set as Adam optimizer,³⁶ learning rate 1 × 10⁻⁴, weight decay 1 × 10⁻³; epoch amount 120, and learning rate decayed by 10 by every 30 epochs. The results of the comparison are given in Table 8, where the acc0, acc0.5 and acc1 represent the classification accuracy of the method under errors of 0, 0.5 and 1 degree respectively, which verifies that the proposed ensemble classifier achieves superior classification accuracies compared with other single feature sets or the existing methods.

Table 8.

Results obtained by existing models

Method	acc0 (%)	acc0.5 (%)	acc1 (%)	Execution time (sec. per sample)
$F_{edge}^{1}$ +kNN^15, ¹⁶	76.88	95.58	99.74	1.339
$F_{glcm}$ +BP¹³	68.83	92.21	99.74	2.165
$F_{glcm}$ +SVM¹⁴	77.40	93.25	100	2.161
$F_{wavelet}^{1}$ +SVM¹⁸	72.21	90.91	99.48	1.096
$F_{fft}^{17}$ +SVM	79.22	94.55	100	1.063
LeNet³²	50.91	73.25	90.91	9.111
AlexNet³³	76.62	93.77	100	8.250
Inception V3³⁴	73.51	93.77	100	26.728
ResNet 50³⁵	77.92	94.81	100	18.267
Proposed method	82.60	97.14	100	3.498

In respect of real-world application, although the proposed method needs more execution time than the other traditional methods, it is still much faster than the deep learning methods. In addition, the execution time of 3.498 seconds per sample and the assessment accuracy of acc0 82.60%, acc0.5 97.14% and acc1 100% are applicable to some extent in real applications in enterprises or testing organizations, based on the authors' consultations with the testing departments of the potential user organizations.

Conclusion

In this study, a 2D-image-based objective system for the assessment of fabric smoothness was established. First, an image acquisition system was built with the capability of controlling the illumination environment. With a number of feature sets extracted from the fabric images in different illumination environments, the ideal environment was obtained by a designed strategy. Finally, by the proposed feature selection and classifier ensemble method, the system generated superior classification results with an 82.86% classification accuracy. The study provides a hybrid solution for the assessment of fabric smoothness based on 2D images, which develops the previous studies in this field into the possibility of further real-world application.

There are some limitations in this study, however. Limited by the characteristics of 2D images, only white fabrics are used and discussed in this study. The adaptation of the system to color fabrics with complex and diverse patterns is limited by the challenge of establishing a universal feature set to distinguish a colored pattern from the pattern caused by the wrinkles. The adaptation of the system to color fabrics will be considered in the authors' future research. In addition, the assessment accuracy could be improved by other models that are better adapted to the human vision, which will also be considered in the authors' further study.

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 research was supported in part by the National Key R&D Program of China under Grant 2017YFB0309200, the National Natural Science Foundation of China under Grant 61802152, the Natural Science Foundation of Jiangsu Province under Grant BK20180602, the China Postdoctoral Science Foundation Funded Project under Grant 2018M640453, the Jiangsu Province Postdoctoral Science Foundation under Grant 2018K037B, the Fundamental Research Funds for the Central Universities under Grant JUSRP11805, the Fundamental Research Funds for the Central Universities under Grant JUSRP51907A.

References

Fan J and Hunter L. Engineering Apparel Fabrics and Garments. 1st ed. Cambridge: Woodhead Publishing, 2009.

AATCC 124-2018. Smoothness Appearance of Fabrics after Repeated Home Laundering. American Association of Textile Chemists and Colorists (AATCC), 2018.

ISO 7768-2009. Textiles — Test Method for Assessing the Smoothness Appearance of Fabrics after Cleansing. International Organization for Standardization(ISO), 2009.

Ding

, et al. Characterization and assessment of fabric smoothness appearance based on sparse coding. Text Res J 2018; 88: 367–378.

Amirbayat

Alagha

. Objective assessment of wrinkle recovery by means of laser triangulation. J Text I 1996; 87: 349–355.

. Fabric wrinkle evaluation using laser triangulation and neural network classifier. Opt Eng 1999; 38: 1688–1693.

Yang

Huang

. Evaluating fabric wrinkle degree with a photometric stereo method. Text Res J 2003; 73: 451–454.

Xin

Yan

. Measuring and modeling 3D wrinkles in fabrics. Text Res J 2002; 72: 863–869.

Ding

Wang

, et al. Feature-based 3D reconstruction of fabric by binocular stereo-vision. J Text I 2016; 107: 12–22.

10.

Yao

. 3-D Surface reconstruction and evaluation of wrinkled fabrics by stereo vision. Text Res J 2009; 79: 36–46.

11.

Reed

. Instrumental evaluation of fabric wrinkle recovery. J Text I 1995; 86: 129–135.

12.

Zaouali

Msahli

El Abed

, et al. Objective evaluation of multidirectional fabric wrinkling using image analysis. J Text I 2007; 98: 443–451.

13.

Mori

Komiyama

. Evaluating wrinkled fabrics with image analysis and neural networks. Text Res J 2002; 72: 417–422.

14.

Wang

Yao

. Grading the crease recovery with twisting of fabric by using image identification technique. Int J Cloth Sci Tech 2000; 12: 114–123. .

15.

Abril HCC, Valencia E and Millán MS. Objective assessment of wrinkled fabrics by optical and digital image processing. AIP Conference Proceedings 2008; 992: 1045–1050.

16.

Abril HCC, Millán MS and Valencia E. Influence of the wrinkle perception with distance in the objective evaluation of fabric smoothness. J Opt a-Pure Appl Op 2008; 10: 1–10.

17.

Choi

Kim

Jin

, et al. Objective wrinkle evaluation system of fabrics based on 2D FFT. Fiber Polym 2009; 10: 260–265.

18.

Sun

Yao

, et al. Fabric wrinkle characterization and classification using modified wavelet coefficients and support-vector-machine classifiers. Text Res J 2011; 81: 902–913.

19.

. A sub-pixel stereo matching algorithm and its applications in fabric imaging. Mach Vision Appl 2009; 20: 261–270.

20.

Kang

Cho

Kim

. New objective evaluation of fabric smoothness appearance. Text Res J 2001; 71: 446–453.

21.

Abidi

Hequet

Turner

, et al. Objective evaluation of durable press treatments and fabric smoothness ratings. Text Res J 2005; 75: 19–29.

22.

Hesarian

. Evaluation of fabric wrinkle by projected profile light line method. J Text I 2010; 101: 463–470.

23.

Goodfellow I, Bengio Y and Courville A. Deep Learning. Cambridge, Massachusetts, USA: MIT Press, 2016.

24.

Shi

Zhang

, et al. Deep CNN with multi-scale rotation invariance features for ship classification. IEEE Access 2018; 6: 38656–38668.

25.

Chen

Sun

Wang

, et al. A novel model based on adaboost and deep CNN for vehicle classification. IEEE Access 2018; 6: 60445–60455.

26.

Lee

Kwon H . Going deeper with contextual CNN for hyperspectral image classification. IEEE T Image Process 2017; 26: 4843–4855.

27.

Zuo

Zhou

, et al. Multi-resolution CNN and knowledge transfer for candidate classification in lung nodule detection. IEEE Access 2019; 7: 32510–32521.

28.

Malek

Drean

Bigue

, et al. Optimization of automated online fabric inspection by fast Fourier transform (FFT) and cross-correlation. Text Res J 2013; 83: 256–268.

29.

Chang

Lin

. LIBSVM: A library for support vector machines. . Acm T Intel Syst Tec 2011, pp. 2.

30.

Zhou

. Machine Learning, 1st ed. Beijing: Tsinghua University Press, 2016.

31.

Paszke A, Gross S, Chintala S, et al. Automatic differentiation in pytorch. NIPS Workshop Autodiff 2017; 1–10.

32.

Lecun

Bottou

Bengio

, et al. Gradient-based learning applied to document recognition. P IEEE 1998; 86: 2278–2324.

33.

Krizhevsky A, Sutskever I and Hinton GE. Imagenet classification with deep convolutional neural networks. NIPS 2012; 1097–1105.

34.

Szegedy C, Vanhoucke V, Ioffe S et al. Rethinking the inception architecture for computer vision. IEEE CVPR 2016; 2818–2826.

35.

He K, Zhang X, Ren S et al. Deep residual learning for image recognition. IEEE CVPR 2016; 770–778.

36.

Kingma DP and Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980 2014.