Image retrieval of wool fabric. Part I: Based on low-level texture features

Abstract

With huge and ever-growing products in the factory, image retrieval can help the worker retrieve the same, or similar, existing products rapidly and accurately to guide production. In this paper, an effective method based on Fourier transform and local binary pattern is proposed to improve the retrieval efficiency of wool fabric. After capturing the fabric image, histogram equalization was implemented on the value of the Hue, Saturation, Value (HSV) mode to enhance the contrast. Subsequently, Fourier transform together with local binary pattern operator were performed to obtain the frequency spectrum and the local binary pattern, respectively. Each frequency spectrum was divided into 22 rings with the same width, and the standard deviation of the frequencies in each ring was calculated as a Fourier feature. Distinct output values of each local binary pattern were counted and normalized as local binary pattern features. Finally, Euclidean distance was adopted to measure the similarity based on the Fourier feature and local binary pattern feature. Twenty thousand wool fabric images were captured to demonstrate the efficacy of the proposed method. Experimental results indicate that the framework is effective and superior for image retrieval of wool fabric, providing referential assistance for the worker in the factory and improving retrieval efficiency.

Keywords

wool fabric Fourier transform local binary pattern image retrieval

With the development of modern multimedia technologies, the number of images in databases has been increasing rapidly. Effective retrieval of the image database is the main problem, especially in wool fabric factories. Traditionally, when a factory gets samples from consumers for imitation, the first step is analyzing the samples and manually searching for the same or similar existing fabrics in the warehouse. Then, the process sheet is obtained to guide production. This method is very time consuming, laborious, and with low retrieval accuracy. At present, image retrieval of wool fabric is achieved by text-based image retrieval (TBIR)¹ in the factory. TBIR relies on textual keywords manually annotated on images, which improves the retrieval efficiency. However, the manual annotation of the images is also very time consuming, tedious, and subjective. Thus, content-based image retrieval (CBIR)² has been developed to overcome the disadvantages inherent in TBIR.

Recently, CBIR has been successfully implemented in different fields, such as medical diagnosis,³ geographic information⁴ and remote sensing,⁵ trademark and intellectual property,⁶ and so on. The methods of CBIR include texture-based,^7,8 color-based,⁹ shape-based,¹⁰ multi features fusion,¹¹ and deep learning.¹² In the textile and garment industry, research into CBIR mainly focuses on patterned fabric, including fabric patterns,^13,14 printed fabric,^15,16 Jacquard fabric,¹⁷ and lace fabric.¹⁸ The adopted methods of texture extraction include unsupervised color image segmentation,¹³ Gabor filters,¹⁵ singular value decomposition,¹⁷ multi-scale and rotation invariant local binary pattern (MRLBP),¹⁸ speeded up robust features (SURF),¹⁴ and edge histogram.¹⁶ These frameworks all obtain satisfactory results but are only suitable for patterned fabric. For fabric image retrieval, texture extraction methods have been performed by Fourier transform (FT)¹⁹ and Gabor filters,^19,20 LBP descriptors,²¹ fractal-based texture,²² scale-invariant feature transform (SIFT),²³ and deep learning.²⁴

To meet the factory requirements, the image retrieval of wool fabric includes two parts: texture and color. This paper is aimed at the first part, retrieving images based on texture features. Deep learning will not be considered at first due to the large number of training samples required and long training cycle. The adopted woven wool fabrics are divided into plain color, striped, plaid, and patterned fabric in the factory. Wool fabric texture includes the fabric weave at the bottom, and strip, lattice, or pattern on the surface. The fabric weave, which reflects many parameters of the process sheet, has not been considered in previous studies. The fabric weave, strip, and lattice are commonly periodic. FT is suitable for extracting periodic features, but how to reflect the differences is the main problem. The FT feature proposed by Chen et al.¹⁹ only have two values, making different wool fabrics difficult to distinguish. Moreover, some patterns in patterned fabrics are aperiodic and the surfaces of wool fabrics are dark and blurring. A single texture feature cannot characterize complicated wool fabrics.

In this paper, the regional difference of Fourier spectrums was proposed to characterize different periodic fabrics. Different LBP operators were optimized and combined with FT to improve the retrieval precision. This paper proceeds as follows. The second section introduces the proposed methods. The third section describes the experiment implementation. The fourth section presents the experimental results, the results analysis, and discussion of the methods. The paper concludes in the fifth section.

Methods

Image capture and pre-processing

All wool fabrics were collected from the factory. The images were captured in red, green, blue (RGB) mode using a scanner (Canon 9000F Mark II). The light source of the scanner was a white light-emitting diode (LED), which can guarantee a stable capture environment. Image resolution influences the calculation quantity and image quality. From preliminary experiments, the resolution was set to 200 dpi in this experiment. Initially, scale calibration was conducted with a ruler. As shown in Figure 1, there are 396 pixels in a range of 5 cm, i.e. the relationship is 79.2 pixels/cm. Considering the size of the real fabric, all images were captured and cropped to a size of 512 pixels × 512 pixels to remove irrelevant information; the corresponding physical size is 6.46 cm × 6.46 cm. Figure 2 gives four kinds of images after capturing.

Figure 1.

Sketch map of scale calibration.

Figure 2.

The images after cropping.

For further analysis and processing, image enhancement²⁵ was employed to emphasize the edge, contour, and contrast of the wool fabric images. The image was converted to Hue, Saturation, Value (HSV) mode, where the V component describes the color's lightness. Then, histogram equalization was implemented on the V component, and the image was converted to RGB mode. As shown in Figure 3, the enhanced image is bright, and the edge becomes clear. After enhancing, the histogram of the V component becomes sharper. Lightness enhancement strengthens the contrast of the image, being propitious for the extraction of texture features. Figure 4 exhibits the enhanced images of the different wool fabrics shown in Figure 2.

Figure 3.

The comparison of image enhancement.

Figure 4.

The enhanced images of the different wool fabrics.

Fourier transform

FT²⁶ converts an image between spatial and frequency domains using numerical analysis. The frequency spectrum obtained after transforming exhibits the periodic texture information of the image in the spatial domain. Discrete Fourier Transform (DFT) converts a finite sequence of equally spaced samples of a function into the list of coefficients of a finite combination of complex sinusoids, ordered by frequencies. For an image of size M × N, the two-dimensional DFT of discrete signal f(x, y) is defined as²⁷

F (u, v) = \frac{1}{MN} \sum_{x = 0}^{M - 1} \sum_{y = 0}^{N - 1} f (x, y) e^{- j 2 π (ux / M + vy / N)}

(1)

Here, x and y are spatial variables in image space, u and v are frequency variables in DFT domain. Let R(u, v) and I(u, v) be the real part and imaginary part of F(u, v), the Fourier spectrum is defined as

| F (u, v) | = [R (u, v) 2 + I (u, v) 2] \frac{1}{2}

(2)

R (u, v) = \frac{1}{MN} \sum_{x = 0}^{M - 1} \sum_{y = 0}^{N - 1} f (x, y) cos [2 π (ux / M + vy / N)]

(3)

I (u, v) = \frac{1}{MN} \sum_{x = 0}^{M - 1} \sum_{y = 0}^{N - 1} f (x, y) sin [2 π (ux / M + vy / N)]

(4)

The power spectrum is defined by

P (u, v) = log (1 + | F (u, v) |)

(5)

The Fourier feature is represented using DFT in this paper. After converting the enhanced image to a grayscale image, the image was transformed to obtain the Fourier frequency spectrum using DFT. Figure 5(a1) to (a4) give the frequency spectrums of the images in Figure 4. Different textures were transformed to different frequency spectrums. To distinguish different fabrics, each spectrum was divided into several rings to reflect the regional difference. The standard deviations of frequencies in each ring were calculated as Fourier feature to reflect the regional difference, and without geometric invariance. The number of rings has a great impact on the performance of texture feature, and directly influences the retrieval results. Through experiments, the number of rings was set to 22 in the same width d, as shown in Figure 5(b1) to (b4). Figure 5(c) exhibits the features of different fabrics varied in different rings. It proves that Fourier feature can be selected as one feature for image retrieval of wool fabric.

Figure 5.

Different wool fabric images and corresponding Fourier spectrums. (a) Fourier frequency spectrum. (b) Divided areas. (c) Texture features in different rings.

Local binary pattern

LBP²⁸ labels the pixels of an image by thresholding its neighbors with the center pixel and representing the result as a binary string or a decimal number. Commonly, LBP operators include grayscale invariance LBP ( $LB P_{P, R}$ ), rotation invariant LBP ( ${LBP}_{P, R}^{ri}$ ) and grayscale and rotation invariant LBP ( ${LBP}_{P, R}^{riu 2}$ ). $LB P_{P, R}$ is defined in equation (6), where g_c and g_p are the gray value of the center pixel and its neighbors, respectively. P is the number of neighbors, reflecting the scales of the two-dimensional space. R is the radius of the neighborhood, reflecting the resolution of the angle space. Assigning a binomial factor 2^p for each sign s(g_p−g_c), in the spatial structure of the local image texture, each pixel is represented by the decimal number after transforming, as shown in Figure 6.

\begin{matrix} LB P_{P, R} = \sum_{p = 0}^{P - 1} 2^{p} \times s (g_{p} - g_{c}), s (x) = {\begin{matrix} 1, x \geq 0 \\ 0, x < 0 \end{matrix} \end{matrix}

(6)

Figure 6.

Calculation of LBP operator.

${LBP}_{P, R}^{ri}$ is defined as equation (7), and ${LBP}_{P, R}^{riu 2}$ is defined as equations (8) and (9).²⁹ The principle of ${LBP}_{P, R}^{ri}$ is rotating the original pattern to get a series of LBP values, and the pattern with the lowest LBP is selected as the final pattern. For ${LBP}_{P, R}^{riu 2}$ , the numbers of spatial transitions (bitwise 0/1 changes) in the pattern less than 2 are classified as a single class category, respectively, and the numbers more than 4 are classified as a class category. Figure 7 illustrates three circularly symmetric neighboring sets with different P and R.

{LBP}_{P, R}^{ri} = min {ROR (LB P_{P, R}, i) | i = 0, 1, \dots, P - 1}

(7)

Figure 7.

Circularly symmetric neighboring sets with different P and R.

where ROR(x, i) performs a circular bit-wise right shift on the P-bit number x i times.

\begin{matrix} {LBP}_{P, R}^{riu 2} = {\begin{matrix} \sum_{p = 0}^{P - 1} s (g_{p} - g_{c}) if U (LB P_{P, R}) \leq 2 \\ P + 1 otherwise \end{matrix} \end{matrix}

(8)

\begin{matrix} U (LB P_{P, R}) = | s (g_{P - 1} - g_{c}) - s (g_{0} - g_{c}) | \\ + \sum_{p = 1}^{P - 1} | s (g_{p} - g_{c}) - s (g_{p - 1} - g_{c}) | \end{matrix}

(9)

In this paper, with different P and R, three LBP operators were performed on the fabric image to obtain corresponding patterns. Then, occurrence frequencies of distinct output values of each pattern were counted and normalized to form the statistical histogram³⁰ as LBP feature. Through experiments, $LB P_{12, 1.5}$ performed best and was selected to extract the LBP feature. In this pattern, the number of distinct output values is 135, i.e., the length of the LBP feature is 135.

Similarity measure

Commonly, similarity measure methods consist of the Minkowsky distance, Chebyshev distance, quadratic form distance, Euclidean distance and correlation coefficient, and so on.³¹ In this paper, the Pearson correlation coefficient³² and Euclidean distance,^14,15,19 commonly used in image retrieval systems, were tested for similarity measure. In n-dimensional space, assuming that X = (x₁, x₂, …, x_n) and Y = (y₁, y₂, …, y_n) are two vectors, the Pearson correlation coefficient r(X, Y) and Euclidean distance ρ(X, Y) are defined as formulas (10) and (11), respectively. When the r(X, Y) is closer to 1, the two images are more relevant. But when the ρ(X, Y) is closer to 0, the two images are more relevant.

r (X, Y) = \frac{Cov (X, Y)}{\sqrt{Var [X] Var [Y]}}

(10)

where, Cov(X, Y) is the covariance of X and Y. Var[X] is the variance of X, and Var[Y] is the variance of Y.

ρ (X, Y) = \sqrt{\sum_{i = 1}^{n} (x_{i} - y_{i}) 2} i = 1, 2, 3, \dots, n

(11)

Euclidean distance, which performed better than Pearson correlation coefficient through experiments, was chosen for the similarity measure. As to the two distances based on the Fourier feature and LBP feature do not have the same order of magnitude, weight allocation³³ is not suitable for combination in this experiment. Let ρ₁ and ρ₂ represent the distances of the Fourier feature and the LBP feature, respectively; the product of ρ₁ and ρ₂ was used as the integrated distance, as defined in equation (12).

d = ρ_{1} \times ρ_{2}

(12)

Experiment implementation

For image retrieval of wool fabric, there are no public datasets. Specifically, 20,000 wool fabric images were collected from the factory and captured to build the image database. Experiments are conducted in a MATLAB compiling environment on a desktop computer with Intel 3.40 GHz processor and 4 GB RAM. Figure 8 exhibits the examples in the image database.

Figure 8.

Wool fabric samples of different categories.

The procedure of the proposed method is illustrated in Figure 9. The texture features were extracted using FT and LBP operator. The distance between the features of the query image and database image was calculated for similarity measure. The top-K images sorted by the distance from small to large were output as the retrieval results. In this experiment, considering the number of relevant images in the image database, K was set to 8.

Figure 9.

Flow chart of the proposed method.

The retrieval interface was built using MATLAB GUI in this experiment. As shown in Figure 10, the items of the interface are retrieval options, methods of similarity measure, comparison methods, retrieval results, and elapsed time. The experiments were conducted using different options.

Figure 10.

The retrieval interface.

Precision and recall were used to measure the performance of this retrieval system. Precision refers to the ratio of relevant retrieved images to the total retrieved images. Recall refers to the ratio of the retrieved relevant images to the total relevant images. The retrieval results were evaluated by experts to establish what images are relevant. Recall and precision² are represented by equations (13) and (14). S is the set of all relevant images in the image database, and R is the set of all retrieved images by the system. s is the relevant images in set R. u is the irrelevant images in set R, and v is the relevant images but not in the set R.

reccall = P (R | S) = \frac{P (S \cap R)}{P (S)} = \frac{s}{s + v}

(13)

precision = P (S | R) = \frac{P (S \cap R)}{P (R)} = \frac{s}{s + u}

(14)

Results and discussion

This section gives the experimental details and results of parameter optimization, the proposed scheme, and methods comparison. A total of 400 query images, 100 of each category, were selected to perform each experiment. Average precision and average recall were calculated to evaluate the performance.

Experiments for parameter optimization

For the number of rings, undersize or oversize rings will cause the features to be difficult to distinguish. With different numbers of rings, the feature databases were built for experiments. The trial test indicated that the number of rings should be more than 16. Therefore, the number of rings was set to d, d = 16, 17, 18, …, n, until the optimal number of rings was found. There are four cases; namely, precision and correlation coefficient (P-C), recall and correlation coefficient (R-C), precision and Euclidean distance (P-D), and recall and Euclidean distance (R-D). Figure 11 represents the average precision and average recall with different numbers of rings. The number of rings was set to 22, and Euclidean distance was selected for similarity measure, because the precision and recall are the highest under this circumstance.

Figure 11.

Average precision and average recall with different numbers of rings.

For different LBP operators, P determines the number of sampling points, directly influencing the retrieval time. The trial test was implemented to compare the calculation time, and the following LBP operators,

LB P_{8, 1}

LB P_{12, 1.5}

{LBP}_{8, 1}^{ri}

{LBP}_{8, 1}^{riu 2}

and

{LBP}_{12, 1.5}^{riu 2}

were selected for experiments. Table 1 shows retrieval results of different LBP operators. C represents the Pearson correlation coefficient and D represents the Euclidean distance. As seen in Table 1, none of the retrieval times exceed 1.0 second, which is acceptable for image retrieval. Except the retrieval time,

LB P_{12, 1.5}

performed best, so

LB P_{12, 1.5}

was selected to extract LBP feature and the Euclidean distance was chosen for similarity measure.

Table 1.

Average precision, recall, and time of different types of LBP operators

		Precision (%)	Recall (%)	Time (s)
$LB P_{8, 1}$	C	74	47	0.358
$LB P_{8, 1}$	D	74	48	0.512
$LB P_{12, 1.5}$	C	79	50	0.764
$LB P_{12, 1.5}$	D	81	54	0.907
${LBP}_{8, 1}^{ri}$	C	75	49	0.484
${LBP}_{8, 1}^{ri}$	D	76	49	0.635
${LBP}_{8, 1}^{riu 2}$	C	70	44	0.448
${LBP}_{8, 1}^{riu 2}$	D	71	44	0.598
${LBP}_{12, 1.5}^{riu 2}$	C	77	50	0.762
${LBP}_{12, 1.5}^{riu 2}$	D	79	50	0.926

LBP: local binary pattern.

Experimental retrieval results

The aforementioned methods were performed for experiments. To give a qualitative analysis of our method, Figure 12 presents the top eight retrieved images of different kinds. As seen from Figure 12, the proposed method can successfully retrieve images which have the same or similar textures. Most images of the same texture rank top, and other retrieved ones also have a similar textural pattern, such as the seventh image of Figure 12(a) and the sixth image of Figure 12(d).

Figure 12.

Retrieval results of randomly selected query images.

The average precision and recall of four kinds of wool fabrics are listed in Table 4. From Table 4, the average precision of the four categories is 85% and the average recall is 54%. This average precision means that an average 6.8 images of the eight images were relevant, and the remaining irrelevant 1.2 images were ranked in the last three positions. This system aims at providing references for the worker during fabric retrieval, and the images ranked in the higher positons will be considered first. Therefore, the results were affirmed to be acceptable by the experts in the factory. Moreover, the average retrieval time is 1.4 seconds for 20,000 images, which is faster than the TBIR currently adopted in the factory. The results indicate that the proposed method is effective for image retrieval of wool fabric, which can provide assistance for the worker in the factory and improve the efficiency of fabric retrieval.

Wool fabric classification aims at calculating the precision and recall of different kinds of fabrics, then analyzing the performances of the proposed method. From Table 4, the average precision of striped fabric is the lowest of the four categories, and the results of the other categories are higher than 85%. Through analysis, for plain color, plaid, and patterned fabric, the periodic feature are strong and the boundaries of different regions are evident. Thus, Fourier feature and LBP feature both performed well. Striped fabric indicates that there are strips in the fabric. But when the strips are indistinct, the retrieval results are easily influenced by plain color fabric with the same fabric weave. As exhibited in Figure 13, the strips of (a) and (c) are similar, but the fabric weaves are dissimilar. The fabric weave of (a) and (b) is the same, which is the main texture feature of the fabric. As listed in Table 2, the similarity of (a) and (b) is greater than (a) and (c). The reason is that the period of the fabric weave is a multiple of the period of the strip, thus the Fourier feature performed badly and the precision decreased for the striped fabric. Although the proposed method performed badly in this case, the retrieval results are also satisfactory when the fabric weave is similar.

Figure 13.

Similar striped fabrics and plain color fabric. (a) Striped fabric. (b) Plain color fabric. (c) Similar striped fabric to (a).

Table 2.

The distances of the images in Figure 13

Distances	Features
Distances	Fourier	LBP	Fourier and LBP
(a) and (b)	0.1159	0.0451	0.0052
(a) and (c)	0.4427	0.2207	0.0977

LBP: local binary pattern.

The analysis was given to explain why the product of two distances was adopted in this experiment. Taking the retrieval result of Figure 12(d) as an example, Figure 14(a) and (b) exhibit the retrieval results separately using the Fourier feature and the LBP feature. In Figure 14(a), the identifiers of eight images are 3064, 3058, 3062, 3071, 3067, 2204, 3987, and 3989. In Figure 14(b), the identifiers of eight images are 3064, 3062, 3058, 3060, 3067, 3071, 3063, and 3485. As shown by the distances listed in Table 3 and the final result shown in Figure 12(d), the images were preserved only when both features performed well. The results indicate that the product of ρ₁ and ρ₂ combines the advantages of the Fourier feature and LBP feature, improving the precision of the retrieval results. As for weight allocation, the orders of magnitude of the two distances are different, 10⁻¹ or 10⁻² for the Fourier feature and 10⁻² for the LBP feature. Moreover, the two features perform differently for different fabrics. Thus, it is difficult to set and optimize the weight of weight allocation.

Figure 14.

Retrieval results for the two features. (a) Fourier feature. (b) Local binary pattern (LBP) feature.

Table 3.

Distances between the query image and the retrieved images

Distances	Identifiers
Distances	3064	3058	3062	3071	3067	3060	3063	2204	3987	3989	3485
ρ ₁	0	0.1117	0.1464	0.2435	0.2705	0.4769	0.5304	0.3048	0.3905	0.4053	0.7572
ρ ₂	0	0.0261	0.0148	0.0521	0.0512	0.0501	0.0579	0.0890	0.1212	0.1155	0.0762
ρ₁ × ρ₂	0	0.0029	0.0022	0.0127	0.0139	0.0239	0.0271	0.0271	0.0473	0.0468	0.0577

Table 4.

Average precision and average recall of different methods

Methods	Plain color		Striped		Plaid		Patterned		Average
Methods	P	R	P	R	P	R	P	R	P	R
Gabor filters	73	44	62	43	76	45	79	55	72	47
MRLBP	85	51	82	57	81	48	73	44	80	50
SURF	70	42	57	40	72	43	73	44	68	42
Wavelet analysis	66	40	57	39	76	45	74	51	68	44
Proposed method	87	54	82	55	85	54	86	55	85	54

P: precision; R: recall; MRLBP: multi-scale and rotation invariant local binary pattern; SURF: speeded up robust features.

Comparison results

To prove the superiority of the proposed method, Gabor filters,^15,19 MRLBP,¹⁸ SURF,¹⁴ and wavelet analysis^34,35 were performed for comparison. As described in the literature,¹⁵ the dimensions were 512 × 512 × 32, resulting in excessive computing. Thus, the mean value¹⁹ of each grid was calculated as texture features, and the dimension was reduced to 512. Kang et al. (2015)¹⁴ removed the texture information of the fabric before extracting features. But, as the texture information is important in this experiment, this pre-processing was not conducted. For MRLBP, there were three compression scales (scale 2, 3, and 4) and three sampling radiuses (R = 2, 5, and 8), and the sampling number P was 8. For wavelet analysis, each texture image was characterized using the variances of three components obtained by decomposing at I frequency layers, where I ranged from 1 to 5. Different methods were tested using the mentioned database and Euclidean distance. Table 4 lists the retrieval results for different methods. P represents precision and R represents recall. From Table 4, Gabor filters perform badly on striped fabric. The MRLBP is not suitable for patterned fabric. As the texture information cannot be removed in this experiment, SURF performs badly. Wavelet analysis does not reflect the density and the position of the fabric texture, so the results are not satisfactory. The comparison results indicate that the proposed method is superior for image retrieval of wool fabric.

Conclusion

Image retrieval based on FT and LBP is proposed in this paper to improve retrieval efficiency and accuracy for wool fabric. The advantages of the Fourier feature and LBP feature are combined by using the product of two distances. Experimental results show that the average precision is up to 85%, and the average retrieval time is 1.4 seconds. Based on the performance analysis, it is concluded that the proposed scheme is precise and efficient, which performs better than other texture extraction methods. This image retrieval system can help the worker in the factory retrieve the same or similar products precisely and fast, saving lots of labor and material resources.

In future work, improvements will be made to improve the precision for striped fabric. Moreover, color features will be adopted to meet the demands of the factory, retrieving images which have the same or similar color. Meanwhile, color features will be combined with the texture features to develop a more robust and higher precision image retrieval system for wool fabric.

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: Postgraduate Research & Practice Innovation Program of Jiangsu Province (No. KYCX18_1819);the Fundamental Research Funds for the Central Universities (No. JUSRP51631A); the National Natural Science Foundation of China (No. 61802152), the Natural Science Foundation of Jiangsu Province (No. BK20180602), the China Postdoctoral Science Foundation Funded Project (No. 2018M640453), the Jiangsu Province Postdoctoral Science Foundation (No. 2018K037B).

References

Farruggia

Magro

Vitabile

. A text based indexing system for mammographic image retrieval and classification. Future Gener Comput Syst 2014; 37: 243–251.

Smeulders

AWM

Worring

Santini

, et al. Content-based image retrieval at the end of the early years. IEEE Trans Pattern Anal Mach Intell 2000; 22: 1349–1380.

Yang

Jin

Mummert

, et al. A boosting framework for visuality-preserving distance metric learning and its application to medical image retrieval. Pattern Anal Mach Intell 2009; 32: 30–44.

Yang

Newsam

. Geographic image retrieval using local invariant features. IEEE Trans Geosci Remote Sens 2013; 51: 818–832.

Ferecatu

Boujemaa

. Interactive remote-sensing image retrieval using active relevance feedback. IEEE Trans Geosci Remote Sens 2007; 45: 818–826.

Anuar

Setchi

Lai

. Trademark image retrieval using an integrated shape descriptor. Expert Syst Appl 2013; 40: 105–121.

Jyothi

Madhaveelatha

Mohan

PGK

, et al. Steerable texture descriptor for an effective content-based medical image retrieval system using PCA. Adv Intell Syst Comput 2016; 379: 289–298.

Irtaza

Jaffar

. Categorical image retrieval through genetically optimized support vector machines and hybrid texture features. Signal Image Video Process 2014; 9: 1–17.

Lyu

Zhang

Zha

, et al. Developing a two-step retrieval method for estimating total suspended solid concentration in Chinese turbid inland lakes using Geostationary Ocean Color Imager (GOCI) imagery. Int J Remote Sens 2015; 36: 1385–1405.

10.

Goyal

Walia

. Variants of dense descriptors and Zernike moments as features for accurate shape-based image retrieval. Signal Image Video Process 2014; 8: 1273–1289.

11.

Singh

Pooja . An effective image retrieval using the fusion of global and local transforms based features. Opt Laser Technol 2012; 44: 2249–2259.

12.

Liu

Guo

, et al. Fusion of deep learning and compressed domain features for content based image retrieval. IEEE Trans Image Process 2017; 12: 5706–5717.

13.

Lachkar

Benslimane

D'Orazio

, et al. A system for textile design patterns retrieval. Part I: Design patterns extraction by adaptive and efficient color image segmentation method. J Text Inst 2006; 97: 301–312.

14.

Kang

Zhang

, et al. Fabric pattern retrieval based on maximum stable extremal regions. J Text Res 2015; 36: 156–160.

15.

Jing

, et al. A new method of printed fabric image retrieval based on color moments and gist feature description. Text Res J 2015; 86: 1137–1150.

16.

Xiang

Qian

, et al. Print fabric pattern retrieval based on edge and color features. J Text Res 2018; 5: 137–143.

17.

Zhang

. Image retrieval of warp knitting fabric based on singular value decomposition. Cases Appl 2015; 34: 90–92.

18.

Zhang L, Liu X, Lu Z, et al. Lace fabric image retrieval based on multi-scale and rotation invariant LBP. In: International Conference on Internet Multimedia Computing and Service, Hunan, China, 19–21 August 2015, pp.1–5. New York: ACM.

19.

Chen

Sobue

Huang

. KANSEI based clothing fabric image retrieval. Lect Notes Comput Sci 2009; 5646: 71–80.

20.

Zand

Doraisamy

Halin

, et al. Texture classification and discrimination for region-based image retrieval. J Visual Commun Image Represent 2015; 26: 305–316.

21.

Qiu

, et al. Application research of local binary pattern in fabric color card image retrieval. Biotechnol: Indian J 2014; 10: 14412–14419.

22.

Suciati N, Herumurti D and Wijaya AY. Fractal-based texture and HSV color features for fabric image retrieval. In: IEEE International Conference Control System, Computing Engineering, George Town, Malaysia, 2729 November 2015, pp.178–182. Washington: IEEE.

23.

Zhang

Liu

Wei

. A fast fabric image matching and retrieval algorithm based on locality-sensitive hashing and visual word. Digital TV Wireless Multimedia Commun 2017; 815: 25–35.

24.

Deng

Wang

, et al. Learning deep similarity models with focus ranking for fabric image retrieval. Image Vision Comput 2017; 70: 11–20.

25.

Cho

Bui

. Fast image enhancement in compressed wavelet domain. Signal Process 2014; 98: 295–307.

26.

Cooley

Lewis

PAW

Welch

, et al. The fast Fourier transform and its applications. IEEE Trans Educ 1988; 12: 27–34.

27.

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.

28.

Xue

Cheng

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

29.

Ojala

Pietikäinen

Mäenpää

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

30.

Jing

Zhang

Kang

, et al. Objective evaluation of fabric pilling based on wavelet transform and the local binary pattern. Text Res J 2012; 82: 1880–1887.

31.

Wang

Yan

Xue

. High speed moire based phase retrieval method for quantitative phase imaging of thin objects without phase unwrapping or aberration compensation. Opt Commun 2016; 359: 272–278.

32.

Xiang

Chen

Liu

. Person re-identification by fuzzy space color histogram. Multimedia Tools Appl 2014; 73: 91–107.

33.

Liu

. Failure mode and effects analysis using intuitionistic fuzzy hybrid weighted Euclidean distance operator. Int J Syst Sci 2014; 45: 2012–2030.

34.

Liapis

Tziritas

. Color and texture image retrieval using chromaticity histograms and wavelet frames. IEEE Trans Multimedia 2004; 6: 676–686.

35.

Balamurugan V and Anandhakumar P. An integrated color and texture feature based framework for content based image retrieval using 2D Wavelet Transform. In: IEEE International Conference on Computing, Communication and Networking, Thomas, USA, 18–20 December 2009, pp.1–16. Washington: IEEE.