A novel hyperspectral imaging and modeling method for the component identification of woven fabrics

Abstract

The component identification of textile materials is critical for quality control and measurement in the textile field. A novel hyperspectral imaging method and the related identification model are proposed to classify single-component textiles. Firstly, the hyperspectral data of the single-component fabrics were processed to conduct dimensionality reduction based on locally linear embedding (LLE), principal component analysis (PCA), and locally preserving projection (LPP) algorithms. Moreover, the original data of 288 wavelengths from 920 nm to 2500 nm were compressed to keep the typical wavelength regions. After that, these data were imported into two classifiers (decision tree classifier and K nearest neighbor (KNN) classifier) for training, and an identification model based on these training data was developed for the sample classification. The experimental results showed that all the samples could be identified correctly by the established identification model. The recognition rate and the stability of the classifier based on LPP model and KNN classification algorithm were proved to have the highest accuracy in our research.

Keywords

hyperspectral imaging spectral analysis component identification dimension reduction classification algorithm

Traditional textile testing for quality control in the textile field is usually based on subjective evaluation. However, it has been proved that it is obviously affected by the subjective behavior of human judges. Therefore, the quality evaluation results from different judges could be quite different. Sometimes mistakes, errors, or low efficiency cannot be avoided when using some traditional testing methods. Fiber identification is critical for textile and garment engineering (manufacturing, fashion design, and quality assurance).¹ In this case, objective, digital, and fast evaluation methods should be developed to improve the accuracy of the fiber or fabric identification.² It has become rather complicated and time consuming to identify textile materials because of the abundance of fabrics.³ Therefore, it is very valuable to find some fast, useful, and nondestructive methods.

In recent years, the use of image processing technology for textile testing has gradually risen and can avoid the influence of subjective factors. As a potential technique, hyperspectral imaging (HSI) technology can be utilized to discriminate mass fibers rapidly and nondestructively. It can achieve the functions that traditional Fourier transform infrared instruments do not have.⁴ Although the application of HSI in the textile industry is not popular at present, it should be effective to apply hyperspectral recognition to textile identification, with the advantage of being nondestructive.⁵ Based on quick and nondestructive measurement of characteristic information from hyperspectral images, it is as an efficient method for fiber identification.⁶ It is very convenient to obtain hyperspectral images without any pretreatment of the samples. In addition, it is possible to do qualitative analysis or even semi-quantitative analysis.^7,8

So far, some researchers have applied spectroscopy to conduct material identification because spectroscopy can provide additional spectral information.^9,10 According to reports, Fourier transform near-infrared (FT-NIR) spectroscopy can achieve more than 98% of the identification accuracy of different types of cotton foreign bodies.^11,12 HSI can provide spatial information in the form at a certain wavelength or any pixel on the image.¹³ Yang et al. proposed two optimal band selection methods,¹⁴ one based on the reflection difference extreme value distribution and the other on the spectral distinguishability. The results showed that the near-infrared (NIR) band 780–1800 nm is preliminarily determined to be the optimal band of the heterogeneous fiber. Wang et al. used NIR spectroscopy to rapidly and qualitatively identify bamboo fibrils, bamboo viscose fibers, and ramie fibers.¹⁵ They showed that bamboo fibrils, bamboo viscose fibers, and ramie fibers could be quickly identified by NIR spectroscopy without destroying the sample. Guo et al. obtained the binary image by principal component analysis (PCA), independent component analysis, dual-band ratio analysis, and a wavelength combining method,¹⁶ and then established the detection algorithm through discriminant analysis. A detection algorithm of the inner layer impurity was established to determine the optimal wavelength set of the impurity pixel segmentation. Xu et al. collected hyperspectral data of seed cotton, foreign fibers, and pseudo-foreign fibers.¹⁷ Feature extraction and selection were carried out by combining appearance ratio and gray level co-occurrence matrix, matching template to remove cottonseed, and then completing seed cotton fiber identification. The experimental results showed that the recognition rate of white polypropylene fiber and white chemical fiber cloth reaches about 90%. Mustafic et al. used an area-of-interest method to extract the average spectrum of winceyette and seven impurities from hyperspectral fluorescence images.¹⁸ The results demonstrate the potential for classifying impurities using strong fluorescence hyperspectral fluorescence imaging systems, such as paper and clear plastics. According to the previous studies, this indicates that hyperspectral images may be available for examining the interior of an object using transmission mode.^19,20 Therefore, it was desirable to take advantage of hyperspectral transmission images to identify the fiber content of textiles.

In this paper, a novel HSI and modeling method for identification is proposed for the detection and analysis of fiber components. Without damaging the fibers, data preprocessing, feature analysis of the hyperspectral image, and identification model were combined to achieve the digital identification of fiber components. The identification procedure included the establishment of the standard fabric sample database, the acquisition of hyperspectral data, the hyperspectral data preprocessing, the hyperspectral feature extraction, the fiber composition discrimination model establishment, and evaluation of the fiber composition. Finally, the identification model of fabric was validated based on experimental results.

Experimental details

Eight types of woven fabric were selected for the hyperspectral image acquirement and component analysis in this research. Eighty hyperspectral images were collected from each kind of woven fabric, and their spectral curves were obtained. In total, 640 spectral curves were used in our research. The samples were divided into calibration sets (60) and validation sets (20). After pretreatment of the original spectral curve, three different dimensionality reduction methods were adopted to reduce the dimensionality of hyperspectral data. Next, the data after dimensionality reduction were imported into two different classifiers for identification. Finally, the results of each group were analyzed. The flowchart of data collection and processing is shown in Figure 1.

Figure 1.

Flowchart of hyperspectral image analysis.

Hyperspectral image acquisition system setup

A complete HSI system consists of hardware and software. The hardware includes the hyperspectral sensor, data acquisition and storage device, data display device, mirror scanner (fixed position scanning imaging), power supply, mounting bracket, etc. The software includes a camera setting module, data acquisition module, spectral band pre-selection tools, setting tools for switching between scan mode and on-board mode, fast viewing of hyperspectral image data and rapid radiation calibration, creation of radiation calibration file tools, and so on.

In this paper, new hyperspectral image acquisition equipment was established to obtain hyperspectral images of single-component fabric samples, as illustrated in Figure 2. There are five components:

Figure 2.

Hyperspectral image acquisition system: (a) the general framework of the system and (b) a picture of the actual object.

(1) A CCD camera of high-resolution is applied to capture the hyperspectral images of single-component woven fabrics.

(2) A hyperspectral imager, which is a “Gaia Sorter” hyperspectral sorting instrument (Beijing ZOLIX Instrument Co. Ltd.). The spectral range of the instrument is set to be 1000–2500 nm, the spectral resolution is 10 nm, and the number of pixels is 320 × 256.

(3) Two standard light sources, which are located on both sides of the sample holder and are arranged symmetrically, were used to provide illumination.

(4) A closed cassette, which is applied to reduce the effects of external light during image acquisition.

(5) A server with software (ENVI Classic 5.3(64-bit) and MATLAB R2016b software), which is applied for processing and analyzing captured images.

Establishment of textile sample database

Eight kinds of common woven fabrics were selected from the fabric sample database available in Shanghai University of Engineering Science. Four of them were made of natural fibers, including cotton, wool, silk, and linen. The others were made of synthetic fibers, including polyester, polyethylene, nylon, and polyvinyl chloride. There were a total of 640 fabric samples; 80 samples were used for the investigation of each fabric. Each fabric was cut into 5 cm × 5 cm pieces as samples; the images of all the samples were acquired and analyzed. Some samples are shown in Figure 3.

Figure 3.

Fabric samples: (a) linen fabrics, (b) cotton fabrics, (c) polyethylene fabrics, (d) wool fabrics, (e) polyvinyl chloride fabrics, and (f) nylon fabrics.

Acquisition of hyperspectral images and hyperspectral image calibration

According to the compositions of the fabrics, the samples were delivered into the hyperspectral sorter to collect hyperspectral images. Before the acquisition of images, the exposure time must be adjusted to 15 ms to ensure that the collected images are clear and bright. Moreover, the electric mobile platform was equipped with a drive line speed of 10 mm/s to avoid image distortion and the standard whiteboard calibration was performed. During the collection, each sample was scanned 10 times using the spectrometer, and then the average value was calculated. Finally, a three-dimensional data block of size 640 × 320×256 was produced.

Hyperspectral images are affected by spectral reflection radiation information and effects generated by external factors, such as system errors of the sensor itself, atmospheric radiation effects, terrain effects, and mixed pixels, can distort the spectral curve. Therefore, the fabric hyperspectral image must be calibrated. After radiometric correction (Figure 4), image mask (Figure 5), and image filtering (Figure 6), it could reduce noise and enhance contrast of hyperspectral images. The subsequent extraction of interest areas could be carried out. The data of real reflectance could be obtained by radiating correction and eliminating interference. With the high purity and smoothness of the fabrics used in this paper, the flat field method is selected for image processing. Through the mask processing, the processed hyperspectral images only retained the effective fabric image area and the black background area with a reflectivity of 0. What's more, the images could be denoised using mean filtering.

Figure 4.

Radiation correction: (a) standard whiteboard ROI area, (b) pre-calibration spectral curve, and (c) corrected spectral curve.

Figure 5.

Process of image masking: (a) original picture before masking, (b) mask file for fabric sample card, (c) image before mask processing, and (d) image after mask processing.

Figure 6.

Image filtering: (a) sample data curve before image filtering and (b) sample data curve after image filtering.

The purpose of relative radiation correction is to eliminate the difference in radiation between images of different phases. The ideal goal is to make the multi-temporal images have the same radiation characteristics, i.e. the same features have the same radiance on multi-source remote sensing images of different phases. At the same time, it should be considered that the target of the ground object may change during the acquisition of remote sensing images in different time phases. In this case, it is desirable that the object objects that have not changed have the same radiance, the changed object targets have different radiances, and the difference in radiance can reflect changes in the object. The images after radiometric correction are shown in Figure 4.

The occlusion of the image is to be processed with the selected image, graphic, or object to control the area or processing of the image processing. The particular image or object used for overlay is called a mask or template. In optical image processing, the mask can be a film, a filter, or the like. In digital image processing, the mask is a two-dimensional matrix array, and sometimes a multi-value image is also used. The images after image masking are as shown in Figure 5.

Due to imperfections in imaging systems, transmission media, and recording devices, digital images are often contaminated with various noises during their formation, transmission, and recording. In addition, noise is introduced into the resulting image when the input image object is not as expected in some parts of the image processing. These noises often appear as an isolated pixel or block of pixels that cause a strong visual effect on the image. In general, the noise signal is uncorrelated with the object to be studied. It appears in the form of useless information, disturbing the observable information of the image. For digital image signals, the noise table is a large or small extreme value. These extreme values act on the true gray value of the image pixels by adding and subtracting, causing bright and dark interference to the image, which greatly reduces the image quality and influences. Subsequent work includes image restoration, segmentation, feature extraction, and image recognition. The images after image filtering are as shown in Figure 6.

Feature data extraction and establishment of the classifier

After calibrating the hyperspectral images, all the fabric image regions of the 80 hyperspectral images collected for each woven fabric were set to the region of interest (ROI). After the mean filtering, the spectra of each fabric in the 920–2528 nm range were obtained. Observing the original spectral data, it was found that it can change in the same trend and carry less spectral information when there was more noise in the 920–1000 nm and 2400–2528 nm ranges. Therefore, before using the continuous projection algorithm to process the original spectral data, the bands in the two regions were eliminated, and the data of the 288 wavelengths were initially compressed to 250 wavelengths. Each kind of woven fabric (80) was randomly divided into calibration sets (60) and validation sets (20) to facilitate the extraction and classification model of the subsequent characteristic wavelength.

Feature data extraction

Hyperspectral image data involve a large amount of information, which has dozens or even hundreds of bands. It is prone to the Hughes phenomenon because of its information-related nature and information redundancy. The Hughes phenomenon means that the classification accuracy increases at first with increase of the spectral bands or dimensions number, but then it decreases rapidly when the bands reach a certain point.^21,22 The actual use of hyperspectral data was limited because of the Hughes phenomenon. There was no doubt that high-dimensional data processing has always been a challenge for hyperspectral image processing and practical application. Dimension reduction to cut down the dimension of the eigenspace, therefore, is effective to deal with such situations.^23,24 The dimension-reduction approach simplifies the model on a smaller data set and provides a familiar and understandable explanation for complex data.²⁵

Dimensionality reduction consists of two main aspects: feature selection and feature extraction.²³ Feature selection selects the appropriate subset of wavelengths from the original band, and feature extraction is used to extract important band information through feature space transformation.²⁶ Depending on various assumptions, many popular feature extraction methods have been proposed, such as PCA,²⁷ kernel PCA (KPCA), locally linear embedding (LLE),²⁸ isometric feature mapping (ISOMAP),²⁹ linear discriminant analysis (LDA),³⁰ neighborhood preserving embedding, locally preserving embedding, locally preserving projection (LPP), and Laplacian eigenmaps (LEs).^31,32 Band selection based methods aim to select a subset of the original spectral bands with some selection benchmarks.³³

Feature extraction could preserve certain significant spectral or physical properties hidden in the HSI data when compared with feature selection. LLE, PCA, and LPP algorithms were chosen as the algorithm for dimensionality reduction in this research.

Principal component analysis (PCA)

PCA is a statistical method of dimension reduction which transforms its original random vector related to its component into a new random vector whose component is irrelevant by means of an orthogonal transformation. It is shown in algebra that the covariance matrix of the original random vector is transformed into a diagonal array. In geometry, the original coordinate system is transformed into a new orthogonal coordinate system so that it can indicate the P orthogonal directions of the most scattered points among all the points. The multidimensional variable system is then reduced to a lower dimensional variable system with a higher accuracy.

PCA is a projection from the high-dimensional data space to the low dimensional data space along the direction of the covariance maximum. The spectra of 640 single-component fabrics were imported to conduct the PCA. The total variance was as shown in Table 1 and Figure 7.

Figure 7.

The total variance of bands.

Table 1.

Total variance interpretation

Composition	Initial eigenvalue			Extract the sum of the loads			The sum of the rotational loads
Composition	Sum	Percentage of variance	Grand total (%)	Sum	Percentage of variance	Grand total (%)	Sum	Percentage of variance	Grand total (%)
1	232.740	89.861	89.861	232.740	89.861	89.861	90.101	34.788	34.788
2	13.000	5.019	94.880	13.000	5.019	94.880	72.439	27.969	62.757
3	6.554	2.531	97.411	6.554	2.531	97.411	55.598	21.606	84.362
4	5.389	2.081	99.492	5.389	2.081	99.492	39.144	15.114	99.476
5	1.078	0.416	99.908	1.078	0.416	99.908	1.119	0.432	99.908
6	0.219	0.085	99.993	/	/	/	/	/	/
7	0.019	0.007	100.000	/	/	/	/	/	/
…	…	…	…

The PCA was used to optimize the characteristic wavelength in this paper. The principle component wavelengths were extracted according to the size of the variance contribution rate. Meanwhile, the major components representing the original information could be determined. Each principal component was composed of a linear combination of the wavelengths in the original data. By comparing the coefficients in the component coefficient matrix, the corresponding wavelength of the maximum coefficient was the best characteristic wavelength. As illustrated in Table 1, the PCA can extract five principal components: PCA1, PCA2, PCA3, PCA4, and PCA5, and the five principal components contain 99.908% of the original data. Among them, the first principal component PCA1 can represent most of the original information of the spectrogram, so the characteristic wavelength is determined according to PCA1. PCA1 is made up of a linear combination of images in 250 bands:

\begin{matrix} PCA 1 = ϕ_{1} \times μ_{1} + ϕ_{2} \times μ_{2} + ϕ_{3} \times μ_{3} \\ + ϕ_{4} μ_{4} + \dots ϕ_{250} \times μ_{250} \end{matrix}

(1)

In this linear combination, if the weight coefficient was larger, the contribution of the image of the corresponding wavelength to the principal component PCA1 was larger. By comparing the 250 weight coefficients, the first four larger weights are, respectively: $ϕ_{183}$ , $ϕ_{133}$ , $ϕ_{224}$ , and $ϕ_{76}$ , whose corresponding wavelengths were 1934.8 nm, 1654.7 nm, 2164.5 nm, and 1335.4 nm. These four wavelengths could be characterized by PCA to determine the characteristic wavelengths that carry a large amount of information about the original spectral curve.

Locally linear embedding (LLE)

The LLE algorithm assumes that the data points are linear in the local domain, so any point in the neighborhood can be represented by a locally approximation. The LLE algorithm minimizes the reconstructed cost function and finds the optimal weights. The local neighborhood weights of each point can remain unchanged under multi-scale transformations and change in special circumstances. There is no iterative calculation process of the LLE algorithm, which can reduce the computational complexity greatly. Assume that the sample set consists of N D-dimensional vectors Xi. Each sample point can be represented by a weighted linear combination of its nearest neighbor point, in which the weights reflect the information of the outgoing neighborhood. According to this information, the geometrical properties of the original high dimensional space can be retained in this low-dimensional space. The whole dataset is obtained by repeating operations on the local neighborhood. The key of this method is to use local linear structure to reflect the global nonlinear structure.

As can be seen from Figure 8, the LLE algorithm is divided into three steps. The first step is to seek the process of K nearest neighbors, which uses the same nearest neighbor as the K nearest neighbor (KNN) algorithm. The second step is to find the linear relationship between the K neighbors in the neighborhood for each sample, and obtain the linear relationship weight coefficient W. The third step is to use the weight coefficient to reconstruct the sample data in low dimensions.

Figure 8.

Schematic diagram of LLE algorithm.

The specific process is shown in Table 2.

Table 2.

Locally linear embedding (LLE) algorithm

Input: sample set

D = {x_{1}, x_{2}, \dots, x_{m}}

, nearest neighbor number k, dimensionality after reduction d.

Output: low dimensional sample set matrix d.

Step 1: For

i_{1}

i_{m}

, calculate the k nearest neighbors of

x_{i}

by the European distance as a measure.

Step 2: For

i_{1}

i_{m}

, find the local covariance matrix

Z_{i}

, and find the corresponding weight coefficient vector

W_{i}

Z_{i} = (x_{i} - x_{j})^{T} (x_{i} - x_{j}) W_{i} = \frac{{Z_{i}}^{- 1} I_{k}}{{I_{k}}^{T} {Z_{i}}^{- 1} I_{k}}

Step 3: The weight coefficient vector

W_{i}

is used to form the weight coefficient matrix W, and calculate matrix M.

M = (I - W)^{T} (I - W)

Step 4: Calculate the former

d + 1

eigenvalues of the matrix M and the eigenvectors

{y_{1}, y_{2}, \dots, y_{d + 1}}

corresponding to the

d + 1

eigenvalues.

Step 5: The matrix formed by the second characteristic vector to the d + 1 eigenvector is the output low-dimensional sample set matrix

d = {y_{2}, y_{3}, \dots, y_{d + 1}}

Locally preserving projection (LPP)

The LPP method proposed by the linear approximation of Laplace map (LE) has the advantages of manifold learning method and linear dimension reduction method and is widely used in classification problems such as digital recognition and face recognition. As a linear approximation of Laplacian feature mapping, LPP can better reflect the manifold structure of samples and has been widely used in image retrieval and image restoration. LPP is a linear version of LEs that uses linearity to approximate nonlinear dimensionality reduction methods. LDA and LPP are classical linear methods. LDA pays attention to the separability between image data, while LPP focuses on the local relationship of data.

The LPP method is to construct the relationship between the pairs of samples in the space and maintain the relationship in the projection, while preserving the local neighborhood structure of the sample in the space while reducing the dimension, i.e. in the low dimension. Minimizing the distance-weighted sum of squares between neighbor samples in space can also be understood as avoiding the divergence of the sample set and maintaining the original neighbor structure.

Establishment of the classifier

The conventional role of the classifier is to use a given category of known training data to learn classification rules and classifiers, and then the unknown data classification (or prediction). The classification algorithm is divided into two categories, namely, the methods are based on a probability density and the method based on discriminant function. The classification algorithm based on probability density usually refers to means of the Bayesian theoretical system; it is classified by the knowledge of the potential conditional probability density function. Classification methods based on the discriminant function use the training data to estimate the classification boundary and complete the classification without calculating the probability density function.³⁴

In the classification algorithm based on probability density, there are the well-known Bayesian estimation method and maximum likelihood estimation. These algorithms are parameterized and need to presuppose the distribution model of the class, and then use the training data to adjust the parameters in the probability density. The methods based on the discriminant function assumes that the classification rule is represented by some form of discriminant function, and the training sample can be used to represent the parameters in the calculation function and use the discriminant function to directly classify the test data. In this classifier, there are the well-known perceptron methods, least squares error method, support vector machine method, neural network method, and radial basis function method. Decision tree classifier and KNN classification algorithms were chosen as the algorithm for establishment of the classifier.

Decision tree classifier

In the supervised classification, large training data are very common, and decision trees are widely used. Decision trees, which provide a simple and understandable method to describe the process of decision making based on the past knowledge, are tree structures made up by internal nodes and leaf nodes. Each internal node can be split into one or more children relying on an attribute of the training data for deciding, and each leaf node is associated with a class label or an outcome. Because of its highly transparent characteristics, decision trees are promising tools for designing a parallelized learning algorithm.

The decision tree is a decision analysis method that evaluates the project risk and judges its feasibility by constructing a decision tree to obtain the probability that the expected value of the net present value is greater than or equal to zero, based on the known probability of occurrence of various situations. A graphical method involves intuitive use of probabilistic analysis. Since this decision branch is drawn like a tree, it is called a decision tree. The detailed C4.5 decision trees algorithm process is shown in Table 3.

Table 3.

Decision trees C4.5 algorithm

Input: an attribute-valued dataset D

1: Tree = {}

2: if D is “pure” OR other stopping criteria met then

3: terminate

4: end if

5: for all attributes a ∈ D do

6: Compute information-theoretic criteria if we split on a

7: end for

a_{best}

= best attribute according to above computed criteria

9: Tree = create a decision node that tests

a_{best}

in the root

10: D_V = induced sub-datasets from D based on

a_{best}

11: for all D_V do

12:

Tre e_{v}

= C4.5(D_V)

13: Attach

Tre e_{v}

to the corresponding branch of Tree

14: end for

15: return Tree

KNN classification algorithm

The proximity algorithm, or KNN classification algorithm, is one of the simplest methods in data mining classification technology. The so-called K nearest neighbor is the meaning of k nearest neighbors, saying that each sample can be represented by its nearest k neighbors. The core idea of the KNN algorithm (Figure 9) is that if most of the k most neighboring samples in a feature space belong to a certain category, the sample also belongs to this category and has the characteristics of the samples on this category. Querying set of reference mark patterns could define in a d-dimensional input feature space. The method determines the categories of the sample to be divided according to the categories of the nearest one or several samples. The KNN method relies on the limit theorem in principle. In the category decision, only a small number of adjacent samples are concerned. In KNN, the distance between objects is calculated as the non-similarity index between objects, which is to avoid the matching between objects. The distance generally refers to Euclidean or Manhattan distance here:

d (x, y) = \sqrt{\sum_{k = 1}^{n} (x_{k} - y_{k}) 2} (Euclideandistance)

(2)

d (x, y) = \sqrt{\sum_{k = 1}^{n} | x_{k} - y_{k} |} (Manhattandistance)

(3)

Figure 9.

Schematic diagram of KNN algorithm.

Results and discussion

PCA-KNN recognition model

After dimension reduction by PCA, the four characteristic wavelengths were obtained by analyzing the component coefficient matrix, which can represent a large amount of information of the original spectral curve. The four characteristic wavelengths are 1934.8 nm, 1654.7 nm, 2164.5 nm, and 1335.4 nm. Thus, hyperspectral data were reduced from 250 dimensions to four dimensions. The data of these four characteristic wavelengths are introduced into the KNN classifier as a classification standard. This means the 480 training sets are trained in the KNN classifier by the class label of 1–8. The nearest neighbor number of the classifier is set to 5, and Euclidean distance is used instead of the distance between objects. Finally, the classification of 160 test sets would be identified by the classifier of completion training. The result of this model was as shown in Figure 10.

Figure 10.

Result of PCA-KNN recognition model.

The results show that the recognition rates of seven kinds of single-component textiles are 100%, with only the recognition rate of silk model, which is 90%, having low recognition accuracy. This means that all the cotton, wool, polyethylene, linen, polyester, nylon, and polyvinyl chloride fabrics in the model can be successfully identified by the PCA-KNN model established in this paper. Moreover, the silk fabrics also can be identified by the PCA-KNN model established. All in all, the accuracy of PCA-LNN recognition model is high, even the lowest model recognition rate can reach 90%.

PCA-decision tree recognition model

The imported data for classification of PCA-decision tree recognition model was the same as the PCA-KNN recognition model. The data of the four characteristic wavelengths, 1934.8 nm, 1654.7 nm, 2164.5 nm, and 1335.4 nm, were imported into the decision tree classifier as a classification standard, meaning of that the 480 training sets are trained in the decision tree classifier by the class label of 1–8. This set the percentage of the sample that is misallocated on a node with the parameter inc-node = 1. In this implementation it was assumed that a pattern vector with fewer than 10 unique values (the parameter Nu). Finally, the classification of 160 test sets was identified by the classifier of completion training. The result of this model was as shown in Figure 11.

Figure 11.

Result of PCA-decision tree recognition model.

The results show that the recognition effects of silk, polyvinyl chloride, polyester fabrics, and polyethylene fabrics were perfect because their recognition rate was 100%. Clearly, all the polyester, polyethylene, polyvinyl chloride, and silk fabrics in the model can be successfully identified by the PCA-decision tree model established. The recognition effect of cotton fabrics is only 85%, which is the worst result among eight kinds of fabrics. The recognition rate of the other models (cotton, nylon, and silk fabric) was 95%, indicating that the three fabrics also can be identified by the PCA-decision tree model established in this paper. Compared with the PCA-KNN recognition model, the PCA-decision tree recognition model has lower accuracy.

LLE-KNN recognition model

After dimensional reduction by LLE, the data of 3–15 dimensions were obtained which can represent a large amount of information of the original spectral curve. Thus, hyperspectral data was reduced from 250 dimensions to 3–15 dimensions. The data of these dimensions were imported into the KNN classifier as a classification standard, meaning that the 480 training sets are trained in the KNN classifier by the class label of 1–8. The nearest neighbor number of the classifier was set to 5 and Euclidean distance is used instead of the distance between objects. Finally, the classification of 160 test sets would be identified by the classifier of completion training. The result of this model was as shown in Figure 12.

Figure 12.

Result of LLE-KNN recognition model.

Figure 12 shows that the recognition rates of seven kinds of single-component textiles were 100% while the recognition rate of polyethylene was 90%. There is no doubt that all the cotton, wool, silk, linen, polyester, nylon, and polyvinyl chloride fabrics in the model can be successfully identified by the LLE-KNN model established in this paper. However, the polyethylene fabrics also could be identified by the LLE-KNN model established in this paper. The accuracy of LLE-KNN recognition model established in this paper was excellent, although the lowest model recognition rate also could reach 90%.

LLE-decision tree recognition model

The imported data for classification of the LLE-decision tree recognition model was the same as the LLE-KNN recognition model. The data of these dimensions are imported into the decision tree classifier as a classification standard; the 480 training sets are trained in the decision tree classifier by the class label of 1–8. It sets the percentage of the sample that is misallocated on a node with the parameter inc-node = 1. Through this implementation, the parameter Nu which means unique values is set to be less than 10, contains only 1–10 unique values. Finally, the classification of 160 test sets could be identified by the classifier of completion training. The result of this model was as shown in Figure 13.

Figure 13.

Result of LLE-decision tree recognition model.

According to Figure 13, the accuracy of the recognition model of various fabrics changes with the change of dimension. From the Figure 13, we do not see the cotton fabric and PVC fabric curve because the two kinds of fabric recognition model accuracy rate were 100%. The accuracy of the polyethylene and flax recognition model was maintained at more than 90%, indicating that polyethylene and linen fabrics could be identified by the LLE-decision tree model established in this paper. The recognition model nylon fabric is better than the former two. When the dimension was 11, the minimum recognition rate is 95%. The nylon fabric could be well recognized by the LLE-decision tree model. It was easy to know that recognition effect of polyester, wool, and silk is not great and the LLE-decision tree identification models of three fabrics were not stable. When using appropriate dimensions, its high recognition rate was 100%. When dimensions do not match, the recognition rate was to below 80%. In summary, the LLE-decision tree model was not the best model for hyperspectral recognition of single-component fabrics.

LPP-KNN recognition model

After dimensional reduction by LPP, the data of 3–15 dimensions were obtained which can represent a large amount of information of the original spectral curve. Thus, the hyperspectral data is reduced from 250 dimensions to 3–15 dimensions. The data of these dimensions are imported into the KNN classifier as a classification standard, the 480 training sets were trained in the KNN classifier by the class label of 1–8. The nearest neighbor number of the classifier was set to 5, and Euclidean distance was used instead of the distance between objects. Finally, the classification of 160 test sets was identified by the classifier of completion training. The result of this model was as shown in Figure 14.

Figure 14.

Result of LPP-KNN recognition model.

According to Figure 14, the accuracy of the recognition model of linen fabrics is changing with the change of dimension. When dimension is 5, 6, 8, 9, 10, 11 or 15, the accuracy of just fit was 85%. The recognition effect of linen fabric was not good, and the stability of model recognition was also a poor while the recognition effect of the other fabrics was good enough and the stability of model recognition was acceptable.

LPP-decision tree recognition model

The imported data for classification of the LPP-decision tree recognition model was the same as LLE-KNN recognition model. The data of these dimensions are imported into the decision tree classifier as a classification standard; the 480 training sets were trained in the decision tree classifier by the class label of 1-8. This set the percentage of the sample that is misallocated on a node with the parameter inc-node = 1. In this implementation, the parameter Nu which means unique values is set under 10, contains only 1–10 unique values. Finally, the classification of 160 test sets was identified by the classifier of completion training. The result of this model was as shown in Figure 15.

Figure 15.

Result of LPP-decision tree recognition model.

According to Figure 15, when the dimension is 6 or 8, the worst recognition rate of cotton fabrics is only 85%. When the dimension is 8, the worst recognition rate of wool fabrics was 90% and the worst recognition rate of silk fabrics was 90% when the dimension was 10. In other cases, the accuracy of each model could be above 95%. In summary, the LPP-decision tree model was not the best model for hyperspectral recognition of single component fabrics.

Conclusions

In this paper, eight kinds of common woven textiles were identified by the HSI analysis method. The 80 samples of eight kinds of fabrics collected were divided into verification sets (60) and test sets (20). Three different dimensionality reduction methods (PCA, LLE, and LPP) were used to reduce the dimensionality of the pretreated hyperspectral data. In the form of cross validation, the data after dimension reduction are respectively imported into two different classifiers, decision tree classifier and KNN classifier. Moreover, six identification models were established, which were PCA-KNN, PCA-decision tree, LLE-KNN, LLE-decision tree, LPP-KNN, and LPP-decision tree recognition models. Eight kinds of single component fabrics were identified by using these six models. The experimental results show that these six models can identify eight kinds of single-component textile effectively. Among them, the recognition rate of six models for each fabric is above 80%. When LPP is chosen as dimension reduction method, and KNN is used as classifier, it produces the best recognition effect on hyperspectral images of woven fabric under specific conditions and the precision of the model is the highest relatively. This shows that HSI technology can be applied to the identification of single-component fabrics, including cotton fabrics, polyester fabrics, polyethylene fabrics, wool fabrics, PVC fabrics, nylon fabrics, linen fabrics, and silk fabrics.

Footnotes

Declaration of conflicting interests

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

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Natural Science Foundation of China (61876106) 2017 Talents Action Program of Shanghai University of Engineering Science (Grant No. 2017RC432017) and Shanghai Natural Science Foundation Project (Grant No. 18ZR1416600).

References

Robertson

. Identification of textile fibers. Aust J Forensic Sci 2009; 42: 153–154.

Houck

. Identification of textile fibers, Elsevier, 2009.

E Lidijafras

Simona

Karin

. Characterization of amino groups for cotton fibers coated with chitosan. Text Res J 2009; 79: 219–226.

Abidi

Hequet

. Fourier transform infrared analysis of cotton contamination. Text Res J 2007; 77: 770–778.

Zhan

Yawen

Wang

, et al. Study on the preparation and characteristics of PVDF/CA blend ultrafiltration membrane. Membr Sci Technol 2002; 22(6): 4–8.

Zeng Q, Su G, Long X, et al. Chain, continuous and no-waste method for degumming and fiber-separating the ramie. Patent 9,206,544, USA, 2015.

Houck

The forensic identification of textile fibers. In: Houck

(ed). Identification of textile fibers, Elsevier, 2009, pp. 259–274.

Hayes

P A

Vahur

Leito

. ATR-FTIR spectroscopy and quantitative multivariate analysis of paints and coating materials. Spectrochim Acta Part A 2014; 133: 207–210.

Frederick

Cassill

. Industry clusters and global value chains: analytical frameworks to study the new world of textiles. J Text Inst 2009; 100: 668–681.

10.

Fagan

Defries

Sesnie

, et al. Mapping species composition of forests and tree plantations in northeastern Costa Rica with an integration of hyperspectral and multitemporal Landsat imagery. Remote Sens 2015; 7: 5660–5696.

11.

Chanel

James

Jonn

, et al. Near-infrared classification of cotton lint, botanical and field trash. J Cotton Sci 2012; 16: 72–79.

12.

Fortier

Rodgers

Cintron

, et al. Identification of cotton and cotton trash components by Fourier transform near-infrared spectroscopy. Text Res J 2011; 81: 230–238.

13.

Kasili

Vo-Dinh

. Hyperspectral imaging system using acousto-optic tunable filter for flow cytometry applications. Cytometry Part A 2006; 69: 835–841.

14.

Yang

W Z

D L

Wei

X H

, et al. Selection of optimal band for detecting foreign fibers in lint cotton using spectroscopic analysis. Trans Chin Soc Agric Eng 2009; 25: 186–192.

15.

Wang

Tian

, et al. Structures of bamboo fiber for textiles. Text Res J 2010; 80: 334–343.

16.

Guo J, Rao X, Cheng F, et al. Detection of human and pig hair on the surface of ginned cotton by hyper-spectral imaging. In: 2010 Meeting of the American Society of Agricultural and Biological Engineers, Pittsburgh, PA, 20–23 June 2010.

17.

Wei

Zhou

, et al. Study on identification method of foreign fibers of seed cotton in hyper-spectral images based on minimum noise fraction. IFIP Adv Inf Commun Technol 2016; 392: 166–176.

18.

Mustafic

Jiang

. Cotton contamination detection and classification using hyperspectral fluorescence imaging. Text Res J 2016; 86: 1574–1584.

19.

Kavdir

Buyukcan

, et al. Prediction of olive quality using FT-NIR spectroscopy in reflectance and transmittance modes. Biosyst Eng 2009; 103: 304–312.

20.

Ariana

. Wavebands selection for a hyperspectral reflectance and transmittance imaging system for quality evaluation of pickling cucumbers. Paper 2009; 8: 1–10.

21.

Petropoulos

Kalaitzidis

Prasad Vadrevu

. Support vector machines and object-based classification for obtaining land-use/cover cartography from Hyperion hyperspectral imagery. Comput Geosci 2012; 41: 99–107.

22.

Song

Gong

Zhu

, et al. Wavelength selection and spectral discrimination for paddy rice, with laboratory measurements of hyperspectral leaf reflectance. ISPRS J Photogramm Remote Sens 2011; 66: 672–682.

23.

Sun

Halevy

Benedetto

, et al. UL-Isomap based nonlinear dimensionality reduction for hyperspectral imagery classification. ISPRS J Photogramm Remote Sens 2014; 89: 25–36.

24.

Zhang

, et al. Ensemble manifold regularized sparse low-rank approximation for multiview feature embedding. Pattern Recognit 2015; 48: 3102–3112.

25.

Alimoğlu F and Alpaydin E. Combining multiple representations and classifiers for pen-based handwritten digit recognition. In: Proceedings of the fourth international conference on document analysis and recognition, 18 August 1997, pp. 637-640. New York: IEEE.

26.

Zhao

, et al. Tangent distance-based collaborative representation for hyperspectral image classification. IEEE Geosci Remote Sens Lett 2016; 13: 1236–1240.

27.

Licciardi

Marpu

Chanussot

, et al. Linear versus nonlinear PCA for the classification of hyperspectral data based on the extended morphological profiles. IEEE Geosci Remote Sens Lett 2012; 9: 447–451.

28.

Kim DH and Finkel LH. Hyperspectral image processing using locally linear embedding. In: First international IEEE EMBS conference on neural engineering, 20 March 2003, pp. 316-319. New York: IEEE.

29.

Bachmann

Ainsworth

Fusina

. Exploiting manifold geometry in hyperspectral imagery. IEEE Trans Geosci Remote Sens 2005; 43: 441–454.

30.

Bandos

Bruzzone

Camps-Valls

. Classification of hyperspectral images with regularized linear discriminant analysis. IEEE Trans Geosci Remote Sens 2009; 47: 862–873.

31.

Belkin

Niyogi

. Laplacian eigenmaps and spectral techniques for embedding and clustering. In: Advances in neural information processing systems, MIT Press, 2002, pp. 585–591.

32.

Cahill

Messinger

Czaja

. Schroedinger Eigenmaps with nondiagonal potentials for spatial-spectral clustering of hyperspectral imagery. Proc SPIE 2014; 9088: 908804.

33.

Gao

Zhou

, et al. Determination of featured parameters to cluster stiffness handle of fabrics by the CHES-FY system. Fibers Polym 2013; 14: 1768–1775.

34.

, et al. Characterization of plantar press-comfort performance of warp-knitted spacer fabrics. Text Res J 2018; 88: 357–366.