A new method of printed fabric image retrieval based on color moments and gist feature description

Abstract

Image retrieval has been an active research topic in recent decades. In this paper, a novel and effective algorithm is proposed for printed fabric image retrieval by combining color moments methods and gist feature description methods. Color information distribution could be declared by color moments very well and gist feature description has an advantage in representation of spatial information. Therefore, color features and spatial information features are respectively extracted by color moments and the gist feature description method, which constitutes a feature database. After that, the similarity between query image features and the feature database is computed by Euclidean distance. To demonstrate the efficacy of the framework, experiments are conducted on the FABRIC database. Experimental results indicate that the proposed algorithm is more effective and accurate than other hybrid schemes for printed fabric images, in terms of precision and recall.

Keywords

printed fabric image retrieval color moments gist feature description similarity

Advances in modern multimedia technologies have led to huge and ever-growing achievements of images, audio and video in diverse application areas such as medicine, remote sensing, entertainment, education and on-line information services. This is similar to what occurred in the early computer development stage, during which the amount of alphanumeric data increased rapidly and there were many practical issues in the database system. Effective retrieval of the image database is an essential building block for general multimedia information management. For an image to be searchable, it has to be indexed by its content, which is manually annotated by keywords or automatically extracted by visual features. Although it seems effortless for a human being to pick out flower images from a collection of fabric pictures, object recognition and classification are still the most difficult problems in image identification and computer vision, especially in the textile industry. The study background is to achieve image retrieval for printed fabric images in the factory. In some factories, printed fabric sample images are added to the image database every day. The goal of this paper is to complete image retrieval of printed fabrics with various features.

There has been a rapidly increase in the size of digital printed fabric image collections in recent years. A huge amount of information is stored in these images. It is not possible to access or make use of the information unless it is organized so as to allow efficient browsing, searching and retrieval. There has been very active research in the area of image retrieval since the 1970s, from two major research communities, database management and computer vision.^1–9 To access an images database on content, low-level features are widely used as indexing features for image retrieval to bypass the difficulties of image understanding, such as colors, textures and shapes of objects.

The objective of color and spatial information features is to retrieve all the images whose color composition is similar to the query image. As is known, the color histogram (CH) method is significant to retrieve images, which was first proposed by Swain and Ballard.^10–13 It has been widely used, especially color feature representation of images. However, two disadvantages commonly exist in the CH method, which are respectively noisy interference and lost spatial information. The color histogram and image segmentation (CHIS) method is able to adopt to solve these problems. In image segmenting, the Markov random model (MRF) is chosen to extract spatial information features, which was proposed by Geman and Geman in 1984.^14–16 The MRF model has been applied widely, because that spatial information in an image is described and it has a good theory effectively. In the CHIS method, the printed fabric images are firstly pre-processed with image enhancement, in which red, green, blue (RGB) spaces are transformed into hue saturation value (HSV) spaces to avoid certain concentrated color features of the space.¹⁷ Then the color features and spatial information features are respectively extracted by the CH method and MRF model, which are converted into feature vectors. Similarity between the objective query image and image database is computed by Euclidean distance and displayed in decreasing order. Experimental results show that the CHIS scheme has good precision but low recall for printed fabric images retrieval.

The CH is unable to capture the spatial relationship of color regions. This matter is not only solved by the CHIS scheme, but also dealt with by the color moments (CM) scheme. Many papers and researchers have focused on color indexing approaches based on global distributions of color in an image. The most widely adopted algorithms are the color coherent vector, color correlogram and CM, which are better than the traditional algorithm.^18,19 CM are adapted to extract color features, which are an advantage of a low feature vector dimension and a small number of calculations.^20–24 However, it is inferred that the retrieval efficiency of this method is relatively low, which in practice is often used to filter images. To increase the accuracy of printed fabric image retrieval, an algorithm combined CM and gist feature description (GFD) is presented. In 2001, American scholars Oliva and Torralba were the first to research quick scene identification and classification.^25–27 In this study, filtered images are divided into 4 × 4 equal and non-overlapping grids. Discrete Fourier transform is employed to extract overall spatial information features of printed fabric images, which is estimated by the energy spectrum information of images. Compared with the image segmentation scheme (IS), the GFD scheme could smooth away difficulties with computation and parameter estimation complexity.

The acronyms used in the paper are listed in Table 1.

Table 1.

Glossary

Acronyms	Full expressions
CH	Color histogram
CM	Color moments
IS	Image segmentation
GFD	Gist feature description
CHIS	Color histogram and image segmentation

The structure of the paper is organized as follows.

Feature extraction. One kind of features describes feature extraction, including the color feature and spatial information feature, employing respectively the CM method and GFD method. Another contains a similarity calculation to obtain results of image retrieval.

Similarity matching. Similarity of features is computed by Euclidean distance, in order to measure the standard of retrieval images.

Experimental procedure and experiments. Experimental analysis and discussion are given. This section discusses the experimental results step by step, comparing the results of the proposed algorithm with the CH scheme, CM scheme, IS scheme, GFD scheme and CHIS scheme.

Feature extraction

The key technology of printed fabric image retrieval is looking for certain suitable and efficient features to retrieve. Various feature descriptions for image retrieval have been proposed. From the perspective of visual sense, these could be divided into color features, texture features, shape features and spatial information features. In this paper, color features and spatial information features are adopted to represent overall information of printed fabric images to acquire the computation of similarity between query image and image database.

Color feature

Color is one of the most distinguished and expressive visual features that is utilized in printed fabric image retrieval and object recognition. There are many color descriptors that have been proposed for image retrieval. Lu and Chang²⁸ proposed a new technique for the purpose of more effective printed image retrieval that adopts the color distributions to represent the global characteristics of printed fabric images. CM are known to produce better retrieval accuracy as compared to the conventional color features. In this paper, CM are used for extraction of color features, because primitives of CM are more robust to describe color images.

Image pre-processing

To achieve perfect color features, printed fabric images should be pre-processed. In this paper, image enhancement is used, which means that certain feature of the printed fabric images are emphasized to display, observe or be further analyzed and processed, including edge, contour and contrast. Image enhancement technology is roughly divided into spatial domain and frequency domain. In this study, spatial domain is adopted, meaning that the information of RGB space is converted into the information of HSV space, as shown in Equation (1).²⁹ Certain features of the RGB space are concentrated, which is a disadvantage to extract the color features. However, after converting RGB space into HSV space, color features could be uniformly distributed to prepare for extracting color features. The results of printed fabric image enhancement are viewed by histogram, as shown in Figure 1.

{V = max (R, G, B) S = \frac{max (R, G, B) - min (R, G, B)}{max (R, G, B)} H = {60 \times (G - B) / (S \times V), ifS \neq 0 andmax (R, G, B) = R 60 \times (2 + (B - R) / (S \times V)), ifS \neq 0 andmax (R, G, B) = G 60 \times (4 + (R - G) / (S \times V)), ifS \neq 0 andmax (R, G, B) = B H = H + 360 ifH 〈 0

(1)

where H, S and V represent hue, saturation and value, respectively. R, G and B show the three primary colors of light, red, green and blue.

Figure 1.

Histogram comparing the original image and the enhanced image.

Color moments

CM are simple and effective methods for color feature representation, including mean (first moment), standard deviation (second moment) and skewness (third moment) of colors. Color information is mainly distributed in the low-end moments. Therefore, the first moment, second moment and third moment could express the color distribution of the printed fabric image. These are depicted in Equations (2)–(4):

μ r, j = \frac{1}{N} \sum_{i = 1}^{N} p i, j j = H, S, V

(2)

σ r, j = (\frac{1}{N} \sum_{i = 1}^{N} (p i, j - μ r, j)^{2})^{\frac{1}{2}} j = H, S, V

(3)

s r, j = (\frac{1}{N} \sum_{i = 1}^{N} (p i, j - σ r, j)^{3})^{\frac{1}{3}} j = H, S, V

(4)

where

p i, j

represents the probability of the jth color channel at the ith location in the image. N is the number of pixels in each region. r (

r = {1, 2, 4, 8, 16, 32, 64, 128, 256}

) is the number of non-overlapping sub-images.

Thus the descriptor size for CM is shown as

size (F CM) = r \times n CM \times n C

(5)

where the feature vector

F CM

is given as

F CM = (μ 1, H σ 1, H s 1, H : μ r, H σ r, H s r, H μ 1, S σ 1, S s 1, S : μ r, S σ r, S s r, S μ 1, V σ 1, V s 1, V : μ r, V σ r, V s r, V)

(6)

$n CM$ is the number of CM and n_C is the number of color channels. We compute moments for each color channel separately, in order to increase the retrieval accuracy. Thus, the printed fabric image is divided into r (for example r = 4) non-overlapping sub-images of 64×64 pixels and extracted primitive moments of each color channel from every sub-image, obtaining a descriptor of 12 × 3 floating points.

Spatial feature

GFD is the characteristic of biological inspiration, which imitates human visual features to extract concise contextual information of printed fabric images, and the spatial information feature is also reflected.^30,31 It is introduced briefly as follows.

Gabor filter group

The gist feature is contour information, which is obtained by the multi-direction and multi-scale Gabor filter group. In signal processing, the single-dimensional Gabor function is widely used. In 1985, the two-dimensional Gabor function was extended by Daugman, and then Gabor filter group is formed, as shown in Equation (7):^32,33

g (x, y) = \frac{1}{2 π σ x σ y} \exp [- (\frac{x^{2}}{σ_{x}^{2}} + \frac{y^{2}}{σ_{y}^{2}})] \times \cos (2 π f 0 x + ϕ)

(7)

where x and y are the coordinate location of the image pixel.

σ x

and

σ y

represent, respectively, Gaussian standard deviation in the direction of x and y. f₀ shows the center frequency at the Gabor filter group. ϕ is harmonic factor skewing.

On the basis of the Gabor filter $g (x, y)$ , the multi-direction and a multi-channel filtering scheme is built by scale and rotation transformation of multi-direction and multi-scale Gabor filter group. This is defined as

g sm (x, y) = t^{- s} g (x', y') x' = t^{- s} (x \cos θ + y \cos θ) y' = t^{- s} (- x \sin θ + y \cos θ) m = m π / (m + 1)

(8)

where

t^{- s}

is the scale factor of wavelet expansion. θ is the rotation angle, which is the direction of the Gabor filter. s and m are, respectively, the number of scales and orientations in the Gabor filter group. Different Gabor wavelet filters are obtained by altering the value of s and m.

Gist feature extraction

The principle of feature extraction with the Gabor filter is the convolution between a set of Gabor wavelet functions and printed fabric images. Each Gabor function and its direction of oscillation with a vertical direction produce a strong response. Therefore, significant features of corresponding direction frequency information are detected in the printed fabric images, forming a robust and compact feature representation. A gray-scale image $f (x, y)$ with the size of $c \times d$ is divided into a regular grid with the size of $m p \times m p$ , for which the value is $m G = m p \times m p$ . Each grid is defined as $P i (i = 1, \dots, m G)$ with the size of $c' \times d' (c' = c / m p, d' = d / m p)$ . Meanwhile, the gray-scale image $f (x, y)$ is filtered by the Gabor filter group with s scale and m direction. The Gabor filter of m_d channels performs with convolution ( $m d = s \times m$ ). The result of each grid cascade is defined as the gist feature.

G_{i}^{P} (x, y) = cat m d (f (x, y) * g sm (x, y)) (x, y) \in P i

(9)

where the dimension of

G^{P}

m d \times c' \times d'

G^{P}

is the subset of

G^{I}

(

G^{P} \subseteq G^{I}

), as shown in Equation (10):

G^{I} = {G_{1}^{P}, G_{2}^{P}, \dots, G_{m G}^{P}}

(10)

where the dimension of

G^{I}

m d \times m G

A printed fabric image is divided into 4 × 4 regular grids. Then gist features are acquired with 32 (4 × 8, four scales and eight orientations) Gabor filters. The process of extraction is shown in Figure 2.

Figure 2.

Gist feature extracting process: (a) a patterned fabric image; (b) gist feature description in each grid.

Similarity matching

Printed fabric image similarity typically is assessed by the distance between a set of image features. However, a short distance corresponds to higher similarity and the choice of metrics depends on the type of feature vectors. In order to calculate distance between the query image and image database, Euclidean distance is adopted for the similarity of color and spatial information features.^34,35 We use a weighted color channel scheme, wherein the weight to color moment distance of each color channel is given. Many experimental results show that saturation has more impact in HSV color space; therefore, more weight to the S component (for example w₂) is chosen to be higher than w₁ and w₃, which are both equal. The color moment similarity measurement is defined as follows:

D c (F Q, F ID) = w 1 d^{H} + w 2 d^{S} + w 3 d^{V} {w 1 = 0.25 w 2 = 0.5 w 3 = 0.25

(11)

d^{j} = \sum_{i = 1}^{r} (\sqrt{(μ_{i, j}^{Q} - μ_{i, j}^{ID})^{2}} + \sqrt{(σ_{i, j}^{Q} - σ_{i, j}^{ID})^{2}} + \sqrt{(s_{i, j}^{Q} - s_{i, j}^{ID})^{2}}) j = H, S, V

(12)

s c (F Q, F ID) = \frac{1}{1 + D c (F Q, F ID)}

(13)

where F_Q is the color feature vector of the query image and F_I is the color feature vector of the printed fabric image database. r shows the region of the image.

μ_{i, j}^{Q}, μ_{i, j}^{ID}

are the mean of query and database images, calculated for each color channel i. Analogously,

σ_{i, j}^{Q}, σ_{i, j}^{ID}

s_{i, j}^{Q}, s_{i, j}^{ID}

are respectively standard deviation and skewness of query and database images, belonging to each color channel. All three individual distances

d^{H}, d^{S}, d^{V}

are normalized because of the RGB to HSV model transformation, which causes the value of each constituent component to lie between 0 and 1.

Similarity distance of spatial information described with gist feature descriptors is also computed using Euclidean distance, which is defined as follows:

D G (F Q, F ID) = \sum_{i = 1}^{r} \sqrt{(G_{i}^{Q} - G_{i}^{ID})^{2}}

(14)

s G (F Q, F ID) = \frac{1}{1 + D G (F Q, F ID)}

(15)

When integrated features of printed fabric images are retrieved, the corresponding distance measure is different, due to different extractive image features. Therefore, the similarity distance of sub-features is dealt with by normalization to make different sub-features comparable. The overall similarity is summed to weighted similarity of normalized distance, as shown in Equation (16):

sim (F Q, F ID) = w c s c (F Q, F ID) + w G s G (F Q, F ID)

(16)

where

sim (F Q, F ID)

is the overall similarity of query and database images. w_c,

w G

represent the weighted of color and spatial information similarity, the value of which is respectively assigned as 0.5 with experience.

s c (F Q, F ID) s G (F Q, F ID)

are respectively the similarity of the color feature and the spatial information feature.

Experimental procedure

In this paper, an algorithm combining the CM and GFD schemes is proposed to retrieve printed fabric images effectively, as shown in Figure 3 where the section of feature extraction is shown in Figure 4. The method of printed fabric images retrieval is summarized as follows.

Color feature extraction. Firstly, printed fabric images are pre-processed with image enhancement. Secondly, CM are adopted to extract color features, including mean (μ), standard deviation (σ) and skewness (s) of colors.

Spatial feature extraction. Objective printed fabric images and FABRIC database images are filtered by the Gabor filter group. After that, the gist feature is used to extract spatial features, for which space description is shown in Figure 3.

Similarity matching. The color feature vectors and spatial feature vectors are obtained. Similarity distance of sub-features is processed by normalization to make different sub-features comparable. Similarity of color and space features is respectively computed by Euclidean distance between query printed fabric images and FABRIC database images, which are $s c (F Q, F ID)$ and $s G (F Q, F ID)$ . The overall similarity is summed to weighted similarity of normalized distance, as shown in Figure 4.

Figure 3.

Block diagram of the proposed method.

Figure 4.

Flow charts of mutual information descriptors extraction.

Experimental details

Printed fabric image retrieval database

The proposed algorithm is tested with the FABRIC database containing 700 printed images, which are from the actual factory. Experiments are conducted in a MATLAB compiling environment on a personal computer with Intel 1.60 GHz processor and 1 GB RAM. Fabric printing images with different sizes are contained by a Canon 9000F scanner. The images are divided into seven categories based on their content, namely Flower, Paisley, Stripe, Leaf, Geometry, Spot and Reticulation. Flower with JPEG format is used in a general purpose image database for experimentation. These images are stored and each image is represented by RGB color space. A sample of the FABRIC image database from a textile factory is shown in Figure 5. Precision and recall are calculated to measure retrieval effectiveness for the image retrieval system. Recall measures the ability of the system to retrieve all models that are relevant, which is the ratio of the retrieved relevant images to the total number of relevant retrieved images to the total number of relevant images in the FABRIC database.³⁶ Precision measures the ability of this system to retrieve only models that are relevant, which used the ratio of relevant retrieved images to the total number of relevant retrieved images to the total number of retrieved images. Recall and precision are represented by Equations (17) and (18):

recall = \frac{totalnumberofrelevantretrievedimages}{totalnumberofrelevantimages}

(17)

precision = \frac{totalnumberofrelevantretrievedimages}{totalnumberofretrievedimages}

(18)

Figure 5.

Sample of the FABRIC image database.

Results and analysis

Generally in a printed fabric image retrieval system, images are expressed by multiple visual features. Color information is the most informative feature because of its robustness. Spatial information can be another important feature and previous studies have shown that spatial structure and orientation fit well with the model of human perception.

In this image retrieval, CM are adopted to capture the color content, and GFD is extracted to acquire the spatial information. In order to verify the feasibility of the proposed image features in retrieval, a number of experiments were performed on printed fabric image retrieval, including comparison with the CH scheme, CM scheme, IS scheme, GFD scheme and CHIS scheme. In order to choose the suitable weighting values in integrations with color and spatial features, 300 images are selected as query images from the above image database, and the image retrieval accuracies are tested for different weight values. Figure 6 shows the mean of retrieval precision of 300 query results for different weight values, which reflects the image retrieval accuracy. Figure 7 represents the mean of retrieval recall of 300 query results for different weight values, which reflects the robustness of image retrieval. From Figures 6 and 7, the optimal weight values w_c and $w G$ are calculated as 0.5.

Figure 6.

The mean of retrieval precision of 300 query results for different weight values.

Figure 7.

The mean of retrieval recall of 300 query results for different weight values.

In order to verify the feasibility of the proposed algorithm, the image (Flower) of the FABRIC database is treated as a sample. An example is given in Figure 8 and 9 that was retrieved for a randomly selected image (of Flower) as a query. The user interface of the printed fabric images retrieval system is shown in Figure 8, which is processed in the environment of MATLAB. The image (Paisley) retrieval results of six schemes are shown in Figure 10. Image retrieval by the proposed method is competitive with other methods. The effectiveness of the proposed image retrieval results is from using the CM and GFD schemes. The top 10 images retrieved by the proposed algorithm are depicted in accordance of their rank from left to right following the query printed fabric image.

Figure 8.

User interface of the patterned fabric images retrieval system and result of the proposed scheme.

Figure 9.

The image (Flower) retrieval results using five schemes.

Figure 10.

The image (Paisley) retrieval results using six schemes.

As mentioned, recall and precision are presented to validate the accuracy of the algorithms. In Figure 11, for each kind of printed fabric image, we use 100 images for training. Then 700 images are employed for retrieval for each scheme. Average precision and recall of 100 printed fabric images with each scheme are obtained. The results are shown in Table 2. From Figure 11 and Table 2, the proposed results are compared with other image retrieval approaches. It is inferred from the comparison that the proposed framework acquires good results on the FABRIC database with average precision of 86.3% and average recall of 53.3%. The CHIS method is behind the presented framework in aspects of average precision and recall, which is better than other schemes.

Figure 11.

Average precision and recall of different schemes using 100 images.

Table 2.

Average precision and recall of schemes

Schemes	Average precision (%)	Average recall (%)
Color histogram scheme	55.9	22.3
Color moments scheme	54.9	29.0
Image segmentation scheme	71.0	23.9
Gist feature description scheme	72.0	35.4
Color histogram and image segmentation scheme	81.6	34.7
Proposed scheme	86.3	53.3

To obtain retrieval results for printed fabrics quickly and accurately, the CHIS algorithm compares precision, recall and retrieval time with the proposed framework. Figure 12 shows average precision and recall of different schemes using 700 images with each scheme. Average precision and recall of the proposed scheme are respectively 86.3% and 53.3%. It is shown that the CHIS scheme and the presented scheme are effective to retrieve printed fabric images in terms of average recall and precision. In addition to the comparison given in Table 3, the time taken for printed fabric image retrieval is 0.958 and 9.051 seconds, respectively, for the CHIS scheme and the proposed scheme. This is because 700 images are synchronously retrieved by the CHIS scheme and successively queried by the proposed scheme. Further, we observed that retrieval time of the proposed scheme is more than with the CHIS scheme, but the retrieval performance of our proposed scheme is better than the performance of other schemes in respect of average recall and precision. The results of experiments show that proposed algorithm is accurate and effective.

Figure 12.

Average precision and recall of different schemes using 700 images.

Table 3.

Evaluation parameters of color histogram and image segmentation (CHIS) and proposed schemes

	CHIS scheme			Proposed scheme
	Precision (%)	Recall (%)	Time (s)	Precision (%)	Recall (%)	Time (s)
Flower	81	32	0.963	84	49	9.025
Paisley	79	36	0.957	86	48	9.036
Stripe	88	37	0.932	90	52	9.056
Leaf	82	39	0.948	85	56	9.047
Geometry	78	31	0.978	84	58	9.078
Spot	83	35	0.971	86	53	9.029
Reticulation	80	33	0.957	89	57	9.037
Average	81.6	34.7	0.958	86.3	53.3	9.051

Conclusion

In this paper, a novel framework using a combination of CM and GFD schemes is used in a manner to improve the retrieval accuracy of the printed fabric image retrieval system. CM have a small amount of calculation and gist features and can be stable to capture the representation features of printed fabric images with overall spatial information. Based on the performance analysis, it is concluded that the retrieval accuracy calculated in terms of average precision is 86.3% for the FABRIC database, which is better than other hybrid schemes. Meanwhile, experimental results indicate that the proposed algorithm could retrieve printed fabric images precisely, comprehensively and efficiently.

However, we mainly focus on the building precision and recall rate in the paper instead of the retrieval time. In the future, an optimized algorithm will be presented by combining color, texture and spatial information features of fabric images in order to further reinforce the accuracy of printed image retrieval.

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 received no financial support for the research, authorship, and/or publication of this article.

References

Madhavi

Tamilkodi

Dinakar

. An innovative technique for content based image retrieval using color and texture features. Int J Innov Res Comput Commun Eng 2013; 1: 1257–1263.

Joshi

Bhavsar

Sanghvi

. Image retrieval system using intuitive descriptors. Procedia Technol 2014; 14: 535–542.

Shanmugapriya

Nallusamy

. new content based image retrieval system using GMM and relevance feedback. J Comput Sci 2014; 10: 330–340.

Prabhu

Kumar

. Wavelet based content based image retrieval using color and texture feature extraction by gray level concurrence matrix and color concurrence matrix. J Comput Sci 2014; 10: 15–22.

Zaied

Salwa

Jemai

. A novel approach for face recognition based on fast learning algorithm and wavelet network theory. Int J Wavelets Multiresolution Inf Process 2011; 19: 923–945.

Yan

. Accurate image retrieval algorithm based on color and texture feature. J Multimed 2013; 8: 277–283.

Vogel

Schiele

. Performance evaluation and optimization for content-based image retrieval. Pattern Recogn 2006; 39: 897–909.

Lin

Chiu

Yang

. Finding textures by textual descriptions, visual examples, and relevance feedbacks. Pattern Recogn Lett 2003; 24: 2255–2267.

Sreedhar

Raju

Babu

. Query processing for content based image retrieval. Int J Soft Comput Eng 2011; 1: 122–131.

10.

Wang

Yang

. Robust image retrieval based on color histogram of local feature regions. Multimed Tool Appl 2010; 49: 323–345.

11.

Han

Yang

Lee

. A novel 3-D color histogram equalization method with uniform 1-D gray scale histogram. IEEE Trans Image Process 2010; 20: 506–512.

12.

Suryanto Kim

Kim

. Spatial color histogram based center voting method for subsequent object tracking and segmentation. Image Vis Comput 2011; 29: 850–860.

13.

Xiang

Chen

Liu

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

14.

Shekhovtsov

Kovtun

Hlavac

. Efficient MRF deformation model for non-rigid image matching. Comput Vis Image Understand 2008; 112: 91–99.

15.

Yousefi

Azmi

Zahedi

. Brain tissue segmentation in MR images based on a hybrid of MRF and social algorithms. Med Image Anal 2012; 16: 840–848.

16.

Karadağ

ÖÖ

Yarman Vural

. Image segmentation by fusion of low level and domain specific information via Markov random fields. Pattern Recogn Lett 2014; 46: 75–82.

17.

Cho

Bui

. Fast image enhancement in compressed wavelet domain. Signal Process 2014; 62: 86–93.

18.

Sztandera

Cardello

Winterhalter

. Identification of the most significant comfort factors for textiles from processing mechanical, handfeel, fabric construction, and perceived tactile comfort data. Text Res J 2013; 83: 34–43.

19.

Pojala

Somnath

. Detection of moving objects using multi-channel kernel fuzzy correlogram based background subtraction. IEEE Trans Cybern 2014; 44: 870–881.

20.

Walia

Vesal

Pal

. An effective and fast hybrid framework for color image retrieval. Sens Imag 2014; 15: 1–23.

21.

Wang

Niu

Yang

. A new robust color image watermarking using local quaternion exponent moments. Inform Sci 2014; 277: 731–754.

22.

Francisco

Margarita

Amelia

. Static digits recognition using rotational signatures and Hu moments with a multilayer perception. Engineering 2014; 6: 692–698.

23.

Guo

Dai

Zhu

. Quaternion moment and its invariants for color object classification. Inform Sci 2014; 273: 132–143.

24.

Zhang

Liu

Wang

. Web video thumbnail recommendation with content-aware analysis and query-sensitive matching. Multimed Tool Appl 2014; 73: 547–571.

25.

Wang

Chen

. Rapid identification and classification of Mycobacterium spp. using whole-cell protein barcodes with matrix assisted laser desorption ionization time of flight mass spectrometry in comparison with multigene phylogenetic analysis. Anal Chim Acta 2012; 716: 133–137.

26.

Zhang

Chang

Xiang

. Random forest for intermediate descriptor fusion in shot boundary detection. J Electron 2014; 31: 465–472.

27.

Laurent

. Saliency and gist features for target detection in satellite images. IEEE Trans Image Process 2011; 20: 2017–2029.

28.

T-C

Chang

C-C

. Color image retrieval technique based on color features and image bitmap. Inform Process Manag 2007; 43: 461–472.

29.

Zhao

Chen

Feng

. Infrared image enhancement through saliency feature analysis based on multi-scale decomposition. Infrared Phys Technol 2014; 62: 89–92.

30.

Margherita

Guido

Annalisa

. An overview on molecular biology of KIT/PDGFRA wild type (WT) gastrointestinal stromal tumours (GIST). J Med Genet 2013; 50: 653–661.

31.

José

Alejandro

Guevara-Fujita

. Molecular and genetic features of gastrointestinal stromal tumors (gist). Revista Gastroenterologia Peru 2013; 32: 394–399.

32.

Holighaus

. Structure of nonstationary Gabor frames and their dual systems. Appl Comput Harmon Anal 2014; 37: 442–463.

33.

Wang

. A new Gabor based approach for wood recognition. Neurocomputing 2013; 116: 192–200.

34.

Yuan

. Computing the nearest Euclidean distance matrix with low embedding dimensions. Math Program 2014; 147: 351–389.

35.

Liu

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

36.

Kumar

. A knowledge based reliability engineering approach to manage product safety and recalls. Expert Syst Appl 2014; 41: 5323–5339.