LEDTD: Local edge direction and texture descriptor for face recognition

Signal Processing: Image Communication ∎ (∎∎∎∎) ∎∎∎–∎∎∎

1

Contents lists available at ScienceDirect

3

Signal Processing: Image Communication

5

journal homepage: www.elsevier.com/locate/image

7 9 11

LEDTD: Local edge direction and texture descriptor for face recognition$

13 15 Q1

Jing Li a,b, Nong Sang a,b,n, Changxin Gao a,b

17

a b

National Key Laboratory of Science and Technology on Multi-Spectral Information Processing, Wuhan, China School of Automation, Huazhong University of Science and Technology, Wuhan, China

19 21 23 25

a r t i c l e i n f o

abstract

Article history: Received 22 June 2015 Received in revised form 2 October 2015 Accepted 7 December 2015

A good image representation is critical to face recognition task. Recently, eight-direction Kirsch masks based image descriptors, e.g. local directional pattern (LDP), local sign directional pattern (LSDP), have been devised and shown competitive results than conventional LBP descriptor. However, these methods may lose or do not fully explore valuable texture information of the image. To remedy this drawback, a novel discriminative image descriptor, namely local edge direction and texture descriptor (LEDTD) is proposed in this paper. LEDTD differs from the existing Kirsch based methods in a manner that it not only considers image edge direction information but also extracts image texture feature by encoding the edge response directions of center and its neighborhood pixels by employing local XOR binary coding strategy. Finally, edge direction and texture features are integrated to form the image feature vector. Extensive performance evaluations on four benchmark face databases show that the proposed approach yields a better performance in terms of the recognition rate as well as robustness to the noise compared with the state of the art methods. & 2015 Elsevier B.V. All rights reserved.

27 29 31

Keywords: Face recognition Local binary pattern Kirsch mask Local edge direction Image texture feature

33 35 37

63

39 1. Introduction 41 43 45 47 49 51

Face recognition, as one of the most focused research topic in image processing, pattern recognition and computer vision, has been widely applied in many fields, such as information security, smart cards, entertainment, law enforcement, video surveillance and human–computer interaction. Image feature extraction serves as one of the most critical steps for face recognition. Although numerous approaches have been proposed and tremendous progress has been made, during the past decades, it is still could not perform as well as desired under uncontrolled

53 55 57 59

☆ Fully documented templates are available in the elsarticle package on CTAN. n Corresponding author at: National Key Laboratory of Science and Q2 Technology on Multi-Spectral Information Processing, Wuhan, China. E-mail address: [email protected] (N. Sang).

conditions. Therefore, how to extract discriminative and robust features is of vital importance to face recognition. Generally, the two-dimensional image feature extraction methods in image representation could be broadly summarized into two categories based on their properties, i.e., holistic methods and local methods. The holistic methods generally extract features from a facial image by treating the image as a whole. Principal component analysis (PCA) [1], linear discrimination analysis (LDA) [2], independent component analysis (ICA) [3], locality preserving projection (LPP) [4], local linear embedding (LLE) [5], local discriminant embedding (LDE) [6], marginal Fisher analysis (MFA) [7], discriminant simplex analysis (DSA) [8], nonnegative graph embedding (NGE) [9], clustering-guided sparse structural learning (CGSSL) [10] and robust structured subspace learning (RSSL) [11] are the typical ones of this kind. These methods are liable to be influenced by face image pose, illumination, scale and so on, and variations in these factors can largely degrade

http://dx.doi.org/10.1016/j.image.2015.12.003 0923-5965/& 2015 Elsevier B.V. All rights reserved.

61 Please cite this article as: J. Li, et al., LEDTD: Local edge direction and texture descriptor for face recognition, Signal Processing-Image Communication (2015), http://dx.doi.org/10.1016/j.image.2015.12.003i

65 67 69 71 73 75 77 79 81 83 85

2

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41

J. Li et al. / Signal Processing: Image Communication ∎ (∎∎∎∎) ∎∎∎–∎∎∎

63 edge response into consideration, which means two opposite trends (ascending or descending) of the gradient 65 and contain some more discriminant information. Second, only the most and the second most prominent edge 67 response directions are take into the local pattern coding. Kang et al. [27] proposed the structured local binary kirsch 69 pattern (SLBKP), which quantify the eight edge responses into two four-bit binary codes according to the predefined threshold. Castillo et al. [28] proposed local sign direc- Q3 71 tional pattern (LSDP), similar to the ELDP, the only differ73 ence is that it codes the most and the least prominent edge response directions. Rivera et al. [29,30] proposed local 75 directional texture pattern (LDTP), which is the mixture coding of direction number of local most prominent Kirsch 77 mask edge response and intensity differences along two greatest edge responses directions. 79 In this paper, we propose a novel discriminative and robust image descriptor, local edge direction and texture 81 descriptor (LEDTD), for face recognition. The main novelty of our descriptor can be summarized as follows. (1) Com83 pared with image gray-scale value, the edge direction is more stable than intensity, and the use of edge direction 85 feature makes our descriptor more robust against illumination variations and noise by operating in the gradient 87 domain. (2) Edge responses are not equally important for image feature extraction. We choose directions of the 89 maximum and minimum response, which explicit the gradient direction of bright and dark areas in the neigh91 borhood, to represent local image pixel edge information. 93 (3) Apart from the directional features, local XOR operator is applied to encode image edge direction texture features, 95 which convey power image discriminative information. (4) Our LEDTD makes full use of the center pixel edge 97 direction and surrounding eight neighbor pixels texture information while existing image descriptor LDTP only 99 utilizes four neighbor pixels to encode image local structure. Therefore, LEDTD retains more local structure infor101 mation than the LDTP. Experimental results demonstrate the superiority of our LEDTD compared with the state of 103 the art image representation approaches.

its recognition performance. The local methods usually consider several regions or sets of isolated points, from which features for classification are extracted. Classical methods such as local binary pattern (LBP) [12,13], scaleinvariant feature transform (SIFT) [14,15], speeded-up robust features (SURF) [16], weber local descriptor (WLD) [17], Weber local binary pattern (WLBP) [18], monogenic binary coding (MBC) [19], histograms of local dominant orientation(HLDO) [20], enhanced local directional pattern (ELDP) [21], farthest point distance (FPD) descriptor [22], rotation-invariant fast feature (RIFF) [23], edge orientation difference histogram (EODH) [24] have been widely examined. Compared with holistic methods, local methods are distinctive and invariant to many kinds of geometric and photometric transformations, and have been gaining more and more attention because of their promising performance. Being one of the representative local image descriptors, local binary pattern (LBP) was first introduced by Ojala et al. [12], and it has shown a high discriminative ability for texture classification due to its invariance to monotonic gray level changes. Afterwards, many variants of LBP have been introduced to further improve its performance. However, the feature of all these methods being coded into the bit-string is prone to change due to noise or other variations. Considering that Kirsch compass mask enhances the useful information like edge texture and meanwhile suppresses the external noise effect, recently it has been widely used for image feature extraction. Jabid et al. [25] proposed local directional patterns (LDP), which is an eight-bit binary code calculated by first comparing the absolute edge response values derived from different directional Kirsch masks. Then the top k prominent values are selected and the corresponding directional bits are set to 1, the remaining ð8 kÞ bits are set to 0. Finally, convert the binary number into a decimal one, and the decimal value is the corresponding image pixel LDP expression. Zhong and Zhang [26] proposed the enhanced local directional patterns (ELDP), which improved the LDP in the following two aspects. First, take the sign of the Kirsch

105

43 45

107

Local XOR Coding

109

47

111

49 51

Maxmum Response Direction Image

Maxmum LEDBP Image WPCA

53

61

115 117 119

57 59

113

LEDBP Histogram

Local XOR Coding

55

LEDBP Image Representation

Minimum Response Direction Image

Minimum LEDBP Image

Fig. 1. Flowchart of face image representation using LEDTD.

Please cite this article as: J. Li, et al., LEDTD: Local edge direction and texture descriptor for face recognition, Signal Processing-Image Communication (2015), http://dx.doi.org/10.1016/j.image.2015.12.003i

121 123

J. Li et al. / Signal Processing: Image Communication ∎ (∎∎∎∎) ∎∎∎–∎∎∎

1 3 5

The remainder of the paper is organized as follows. Section 2 presents our LEDTD based image representation in detail and a brief review of SVM based face image classification method. Section 3 conducts experiments to evaluate the performance of the proposed method. Section 4 concludes the paper.

7 9

2. Local edge direction and texture descriptor based image representation

11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55

Our descriptor fully capture the local image edge direction and texture information for feature coding. The overall framework of the proposed approach is illustrated in Fig. 1.

H LEDTD ðs; tÞ ¼

Ri ¼ IM i

i ¼ 0; …; 7

ð1Þ

Second, minimum and maximum edge response direction selection: The Kirsch edge response R0 ; R1 ; …; R7 represent the edge significance in its respective direction, and they are not equally important for image feature extraction. We choose directions of the maximum and minimum response, which explicit the gradient direction of bright and dark areas in the neighborhood, to represent local image pixel edge information. Let (m,n) denotes position of the local central neighborhood pixel, Kmax and Kmin denote the direction number of the maximum and minimum edge response respectively K max ðm; nÞ ¼ arg maxfRk ðm; nÞj0 r k r7g

ð2Þ

K min ðm; nÞ ¼ arg minfRk ðm; nÞj0 rk r 7g

ð3Þ

k

k

Third, local edge direction pattern: The LEDPmin descriptor is derived by applying the XoR operator on the minimum Kirsch edge response direction image. LEDPmin operator describes the minimum Kirsch edge response direction relationships between a pixel and its neighborhood ones. For a given 3  3 pattern, LEDP min value is computed by comparing the center pixel minimum edge response direction number with its neighboring ones as follows: LEDP N;R min ¼

59

( Sðx; yÞ ¼

N 1 X

SðK imin ; K cmin Þ  2i

ð4Þ

i¼0

1

xay

0

x¼y

ð5Þ

O X C X

63 65 67 69 71 73 75

P s ðxÞδðLSPðxÞ ¼ tÞ

ð6Þ 77

o¼1c¼1

1

if A is true

0

otherwise

ð7Þ

79

where O  C is the dimensionality of the image, x is the pixel at location ðo; cÞ in the image coordinates, s ¼ 1; 2; …; S; t ¼ 1; 2; …; T. Note that the size of this 2-D histogram is S  T, where S is the number of edge direction, T is the total number of LEDP patterns. In other words, in this 2-D histogram, each row corresponds to a edge direction ws , and each column corresponds to a LEDP pattern t. Thus, the value of each cell HLEDTD ðs; tÞ corresponds to the frequency of the certain edge direction ws and the LEDP pattern t. After that the 2-D histogram H LEDTD ðs; tÞ is further encoded into 1-D histogram H. We use each row of 2-D histogram to form a 1-D histogram HðsÞ; s ¼ 1; …; S. Each sub-histogram HðsÞ corresponds to the edge direction ws . Concatenating the S sub-histograms, we obtain the 1-D histogram H ¼ H s ; s ¼ 1; …; S. The proposed LEDTD is composed by two description value LEDTDmin and LEDTDmax , image pixel Iðm; nÞ is described by LEDTD as

81

δðAÞ ¼

Fusion of local edge direction and texture information, the proposed local edge direction and texture descriptor (LEDTD) consists of minimum and maximum edge response direction LEDTD, the construction of LEDTD has the following four steps. First, Kirsch edge response computation: Eightdirectional Kirsch masks are shown in Fig. 2. Compute eight directional edge response Ri ; i ¼ 0; …; 7 in the neighborhood by convoluting the image I with the Kirsch masks M i ; i ¼ 0; …; 7 separately as following:

57

61

where N is the number of neighbors, and R is the radius of the neighbor. K imin and K cmin denotes direction number of minimum edge response value of the center and neighbor pixel respectively. The maximum Kirsch edge direction LEDP LEDP max , is computed in the same way, with the difference that we take the maximum direction number in Eq. (4) instead. Fourth, Local edge direction and texture descriptor: LEDTD consists of two components: local edge direction and LEDP, which are both described above. In this work, we first construct the 2-D histogram H LEDTD ðs; tÞ of the original image

 2.1. Local edge direction and texture descriptor

3

LEDTDðm; nÞ ¼ fLEDTDmin ; LEDTDmax g

ð8Þ

83 85 87 89 91 93 95 97 99 101

2.2. LEDTD based face image representation 103 Both the minimum and maximum edge direction LEDTD are 8  256 ¼ 2048 dimensions, we could utilize whitened principle component analysis (WPCA) [31] to reduce the dimensionality. Experimentally, the minimum or maximum edge direction LEDTD dimensionality is reduced to 300. After WPCA, we concatenate the minimum and maximum edge direction LEDTD histograms to form the final feature vector, namely LEDTD histogram, whose dimension is 600. According to above analysis, the minimum and maximum edge direction LEDTD histograms contain complementary information and both have properties desired by image recognition. The concatenation of them should be more discriminative and improve performance theoretically compared with the cases of separately using them. Besides, spatial information is very important for image representation. According to this consideration, we divide the encoded images into m non-overlapping rectangular regions R0 ; R1 ; …; Rm to aggregate the spatial information to the descriptor, and then each of the region is used to

Please cite this article as: J. Li, et al., LEDTD: Local edge direction and texture descriptor for face recognition, Signal Processing-Image Communication (2015), http://dx.doi.org/10.1016/j.image.2015.12.003i

105 107 109 111 113 115 117 119 121 123

J. Li et al. / Signal Processing: Image Communication ∎ (∎∎∎∎) ∎∎∎–∎∎∎

4

1 3 5 7 9 11 13 15 17

Fig. 2. Kirsch edge masks in eight directions.

21

compute a histogram of LEDTD independently. Finally, all image regional LEDTD histograms H 0 ; H 1 ; …; Hm are concatenated into a feature vector as the face image descriptor. Such a descriptor contains information on pixel, regional and global levels.

23

2.3. SVM based face image classification

25

In this work, we adopt support vector machine(SVM), which has been proved to be a powerful classifier for various classification tasks, as classifier to evaluate face recognition performance of different image descriptors. SVM first maps feature data into a higher dimensional space, and then finds the optimal separating hyperplane with maximal margin to split different classes. Given N training samples T ¼ fðxi ; yi Þji ¼ 1; …; Ng, where xi A Rn and yi A f  1; 1g, the test feature data x is classified by function: ! n X αi yi Kðxi ; xÞ þb ð9Þ f ðxÞ ¼ sign

19

27 29 31 33 35 37 39 41 43 45 47 49

i¼1

where n is the number of support vector, αi are Lagrange multipliers of the dual optimization problem, b is the threshold of the hyperplane, and Kð; Þ is a kernel function, in this work, we chose the Radial Basis kernel Function. SVM is binary classifier intrinsically, and the multi-class face recognition classification is achieved by utilizing oneagainst-rest technique, which constructs k classifiers. Gridsearch during 10-fold cross-validation is carried out to select optimal SVM kernel parameters, the parameter setting producing best cross-validation performance is picked. We use the SVM implementation in the publicly available machine learning library LIBSVM [32].

51 53

3. Experiments

55

3.1. Databases and experimental setup

57

The AR face image database [33] contains more than 4000 face images of 126 subjects (70 men and 56 women) with different facial expressions, illumination conditions, and occlusions. For each subject, 26 images were taken in two separate sessions (two weeks interval between the

59 61

two sessions). A subset that contains 100 subjects (50 male and 50 female) is chosen in our experiments and the original images are normalized to 121  100 pixels. The Extended Yale B face database [34] consists of 2414 front-view face images of 38 individuals. There are about 64 images under different laboratory-controlled lighting conditions for each individual, where the horizontal light source angle varies from  130° to þ130°, while the vertical angle changes from  40° to þ 90°. Original image size is 192  168, in our experiments, the cropped images of size 100  100 are used. The CMU PIE face database [35] contains 68 subjects with 41 368 face images as a whole. Images of each person were taken across 13 different poses, under 43 different illumination conditions, and with 4 different expressions. We choose a subset of images from the five near frontal poses (C05, C07, C09, C27, C29) of each person. There are 170 images per subject with all kinds of illuminations and expressions. All images have been cropped and resized to be 64  64 pixels. The FERET face database [36] is a result of the FERET program, which was sponsored by the US Department of Defense through the DARPA Program. It has become a standard database for testing and evaluating state-of-theart face recognition algorithms. The proposed method was tested on a subset of the FERET database. This subset includes 1200 images of 200 individuals (each individual has seven images) and involves variations in facial expression, illumination, and pose. In the experiment, the facial portion of each original image was automatically cropped based on the location of the eyes, and the cropped images was resized to 80  80 pixels. In the experiments, for each database, we randomly selected half of the images from the database of each individual to form the training sample set, the remaining used for testing. Repeated this procedure 10 times and calculated the average recognition rate. For the LEDP descriptor, we use a constant number of 8 sampling directions for radius R ¼ 1, equal to the number of neighbor pixels of original LBP. The number of nonoverlapping subregions, m  m, that a face image is partitioned into, we have set to m ¼ 9.

63 65 67 69 71 73 75 77 79 81 83 85 87 89 91 93 95 97 99 101 103 105

3.2. Performance comparison of LEDTD components

107

Recall from Section 2 that the proposed LEDTD feature is composed of two components, i.e., LEDTDmin and LEDTDmax . In this section, we will perform experiments to investigate the following two issues: (1) Which component contributes more to face recognition performance; (2) The feasibility for the usage of combined LEDTDmin and LEDTDmax features against utilizing them separately. According to Table 1, the following two conclusions can be made. LEDTDmax contributes much more than the LEDTDmin for face recognition, this demonstrates high discriminating capabilities of LEDTDmax features; The combination of both LEDTDmax and LEDTDmin components achieves better results, compared with the cases of separately using them; this implies that LEDTDmax and LEDTDmin components bring different information and

109

Please cite this article as: J. Li, et al., LEDTD: Local edge direction and texture descriptor for face recognition, Signal Processing-Image Communication (2015), http://dx.doi.org/10.1016/j.image.2015.12.003i

111 113 115 117 119 121 123

J. Li et al. / Signal Processing: Image Communication ∎ (∎∎∎∎) ∎∎∎–∎∎∎

1

Table 1

Q4 Recognition rate (%) of different schemes of LEDTD on four databases.

3 5 7

AR Yale B CMU PIE FERET

LEDTDmin

LEDTDmax

LEDTD

86.51 85.29 89.63 90.93

89.16 90.62 90.35 91.69

94.78 93.61 96.60 96.55

9 11

Table 2 Recognition rate (%) comparison with state-of-the-art methods on four databases.

13 15 17 19 21 23

AR

YaleB

CMU PIE

FERET

LBP SLBKP LDP ELDP LSDP LDTP DLMXORP

87.26 87.75 86.92 87.89 88.68 90.31 89.94

85.89 86.62 86.58 87.03 89.97 88.70 88.17

88.39 88.58 89.29 90.84 91.53 92.31 93.83

88.56 88.36 89.58 90.36 91.38 92.70 93.17

LEDTD LEDTD þ WPCA

92.63 94.78

91.96 93.61

94.55 96.60

94.96 96.55

25 27

Table 3 Effects of the additive Gaussian white noise (σ¼ 0.002) on the recognition rate (%).

29 31 33 35 37

AR

YaleB

CMU PIE

FERET

LBP SLBKP LDP ELDP LSDP LDTP DLMXORP

78.61 80.26 78.23 78.73 80.52 83.71 82.36

75.73 79.78 78.63 80.18 82.09 81.66 81.17

79.30 81.18 81.03 83.16 84.09 84.51 86.66

79.42 80.48 81.00 82.81 84.29 84.87 85.54

LEDTD LEDTD þ WPCA

86.48 89.69

85.28 88.52

88.28 91.69

88.68 91.32

39 41

they are complement each other in terms of boosting face recognition performance.

43 3.3. Comparison with other methods 45 47 49 51 53 55 57 59 61

In this section, we will compare the performance of our proposed LEDTD with those of the state-of-the-art local descriptors, including LBP, SLBKP, LDP, ELDP, LSDP, LDTP and DLMXoRP. Table 2 gives the experimental results of the different descriptors. From the results, we can seen that the proposed LEDTD descriptor has higher recognition rate in comparison with all the other descriptor. Meanwhile, it is worth to note that applying WPCA can boost the result on recognition rate of LEDTD, the accuracy of AR database is about 95.98%, which is 2.72% higher than LEDTD. Similarly, for Yale B database, it is 97.06%, while LEDTD performance is about 93.38%. On the CMU PIE and FERET database, by using the WPCA, we can achieve 2.05% and 1.59% improvement in recognition rate over the original feature, respectively. All of these demonstrate the superiority of the proposed LEDTD based

5

Table 4 Effects of the multiplicative speckle noise ðσ ¼ 0:002Þ on the recognition rate (%). AR

YaleB

CMU PIE

FERET

LBP SLBKP LDP ELDP LSDP LDTP DLMXORP

78.83 81.38 78.01 79.23 80.69 82.16 82.95

76.84 79.91 78.33 79.82 81.89 82.51 81.67

78.65 80.69 80.72 83.46 83.67 84.76 86.55

79.00 80.82 80.99 83.15 84.07 85.32 85.95

LEDTD LEDTD þ WPCA

86.92 90.00

85.03 88.16

87.83 90.86

88.11 91.50

63 65 67 69 71 73 75 77

image representation method over other existing ones for face recognition.

79

3.4. Performance comparison under noise

81

For face recognition tasks, noise is inevitable under uncontrolled circumstances. Therefore, it is important for an image descriptor to be insensitive to noise for robust face recognition. In this section, we will verify the tolerance of the proposed method to two common types of image noise, e.g., additive Gaussian white noise and multiplicative speckle noise. The resultant values are listed in Tables 3 and 4. As shown in the tables, the recognition rates of all methods tend to drop in the presence of Gaussian white noise or speckle noise. Among them, the decreases in the recognition rate of our method are least. It can be concluded that our method yields the highest tolerance to the two types of image noise.

83

4. Conclusion In this paper, we have proposed a simple and easy to compute image descriptor, local edge direction and texture descriptor (LEDTD), for face image representation. The main findings of the work are as follows: (1) The proposed LEDTD exploits both image edge direction and texture information available locally, which is evidenced by the improved performance. (2) The WPCA method can further improve the recognition performance of the proposed image descriptor. The experiments have been conducted on four frequently used benchmark face databases to test the generalization performance of our descriptor. The results show that our method has higher recognition rate as well as noise and illumination tolerances compared with the state of the art methods.

85 87 89 91 93 95 97 99 101 103 105 107 109 111 113 115 117

References 119 [1] M. Turk, A. Pentland, Eigenfaces for recognition, J. Cognitive Neurosci. 3 (1) (1991) 71–86. [2] P.N. Belhumeur, J.P. Hespanha, D.J. Kriegman, Eigenfaces vs. Fisherfaces: recognition using class specific linear projection, IEEE Trans. Pattern Anal. Mach. Intell. 19 (7) (1997) 711–720.

Please cite this article as: J. Li, et al., LEDTD: Local edge direction and texture descriptor for face recognition, Signal Processing-Image Communication (2015), http://dx.doi.org/10.1016/j.image.2015.12.003i

121 123

6

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39

J. Li et al. / Signal Processing: Image Communication ∎ (∎∎∎∎) ∎∎∎–∎∎∎

[3] M.S. Bartlett, J.R. Movellan, T.J. Sejnowski, Face recognition by independent component analysis, IEEE Trans. Neural Netw. 13 (6) (2002) 1450–1464. [4] J. Lu, Y. Tan, Regularized locality preserving projections and its extensions for face recognition, IEEE Trans. Syst. Man Cybern. Part B: Cybern. 40 (3) (2010) 958–963. [5] X. Li, S. Lin, S. Yan, D. Xu, Discriminant locally linear embedding with high-order tensor data, IEEE Trans. Syst. Man Cybern. Part B: Cybern. 38 (2) (2008) 342–352. [6] H. Chen, H. Chang, T. Liu, Local discriminant embedding and its variants, In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2005, vol. 2, pp. 846–853. [7] S. Yan, D. Xu, B. Zhang, H. Zhang, Q. Yang, S. Lin, Graph embedding and extensions: a general framework for dimensionality reduction, IEEE Trans. Pattern Anal. Mach. Intell. 29 (1) (2007) 40–51. [8] Y. Fu, S. Yan, T.S. Huang, Classification and feature extraction by simplexization, IEEE Trans. Inf. Forensics Secur. 3 (1) (2008) 91–100. [9] J. Yang, S. Yang, Y. Fu, X. Li, T.S. Huang, Non-negative graph embedding, In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2008, pp. 1–8. [10] Zechao Li, Jing Liu, Yi Yang, Xiaofang Zhou, Hanqing Lu, Clusteringguided sparse structural learning for unsupervised feature selection, IEEE Trans. Knowl. Data Eng. 26 (9) (2014) 2138–2150. [11] Zechao Li, Jing Liu, Jinhui Tang, Hanqing Lu, Robust structured subspace learning for data representation, IEEE Trans. Pattern Anal. Mach. Intell. 37 (10) (2015) 2085–2098. [12] T. Ojala, M. Pietikäinen, T. Mäenpää, Multiresolution gray-scale and rotation invariant texture classification with local binary patterns, IEEE Trans. Pattern Anal. Mach. Intell. 24 (7) (2002) 971–987. [13] T. Ahonen, A. Hadid, M. Pietikanen, Face description with local binary patterns: Application to face recognition, IEEE Trans. Pattern Anal. Mach. Intell. 28 (12) (2006) 2037–2041. [14] C. Fernandez, M.A. Vicente, Face recognition using multiple interest point detectors and sift descriptors, In: Proceedings of the IEEE International Conference on Automatic Face & Gesture Recognition, 2008, pp. 1–7. [15] C. Geng, X. Jiang, Face recognition using sift features, In: Proceedings of the IEEE International Conference on Image Processing, 2009, pp. 3313–3316. [16] D. Philippe, S. Pascal, H. Harald, N. Hermann, Surf-face: face recognition under viewpoint consistency constraints, Proc. Br. Mach. Vis. Conf. (2009) 7.1–7.11. [17] J. Chen, S. Shan, C. He, G. Zhao, M. Pietikäinen, X. Chen, W. Gao, Wld: A robust local image descriptor, IEEE Trans. Pattern Anal. Mach. Intell. 32 (9) (2010) 1705–1720. [18] F. Liu, Z. Tang, J. Tang, Wlbp: Weber local binary pattern for local image description, Neurocomputing 120 (2013) 325–335. [19] M. Yang, L. Zhang, S.C.K. Shiu, D. Zhang, Monogenic binary coding: an efficient local feature extraction approach to face recognition, IEEE Trans. Inf. Forensics Secur. 7 (6) (2012) 1738–1751.

[20] Jianjun Qian, Jian Yang, Guangwei Gao, Discriminative histograms of local dominant orientation (d-hldo) for biometric image feature extraction, Pattern Recognit. 46 (10) (2013) 2724–2739. [21] F. Zhong, J. Zhang, Face recognition with enhanced local directional patterns, Neurocomputing 119 (2013) 375–384. [22] Akrem El-ghazal, Otman Basir, Saeid Belkasim, Farthest point distance: a new shape signature for fourier descriptors, Signal Process.: Image Commun. 24 (7) (2009) 572–586. [23] Gabriel Takacs, Vijay Chandrasekhar, Sam Tsai, David Chen, Radek Grzeszczuk, Bernd Girod, Rotation-invariant fast features for large-scale recognition and real-time tracking, Signal Process.: Image Commun. 28 (4) (2013) 334–344. [24] Xiaolin Tian, Licheng Jiao, Xianlong Liu, Xiaohua Zhang, color-sift: application to image retrieval based on codebook, Signal Process.: Image Commun. 29 (4) (2014) 530–545. [25] T. Jabid, Md.H. Kabir, O. Chae, Facial expression recognition using local directional pattern(LDP), In: Proceedings of IEEE International Conference on Image Processing, 2010, pp. 1605–1608. [26] F. Zhong, J. Zhang, Face recognition with enhanced local directional patterns, Neurocomputing 119 (2013) 375–384. [27] G. Kang, S. Guo, D. Wang, L. Ma, Z. Lu, Image retrieval based on structured local binary kirsch pattern, IEICE Trans. Inf. Syst. E96-D (2013) 1230–1232. [28] J.A.R. Castillo, A.R. Rivera, O. Chae, Facial expression recognition based on local sign directional pattern, In: Proceedings of IEEE International Conference on Image Processing, 2012, pp. 2613–2616. [29] A.R. Rivera, J.R. Castillo, O. Chae, Recognition of face expressions using local principal texture pattern, In: Proceedings of IEEE International Conference on Image Processing, 2012, pp. 2609–2612. [30] A.R. Rivera, J.R. Castillo, O. Chae, Local directional texture pattern image descriptor, Pattern Recognit. Lett. 51 (2014) 94–100. [31] Chengjun Liu, Harry Wechsler, Evolutionary pursuit and its application to face recognition, IEEE Trans. Pattern Anal. Mach. Intell. 22 (6) (2000) 570–582. [32] C.C. Chang, C.J. Lin, LIBSVM: a library for support vector machines, ACM Trans. Intell. Syst. Technol. 2 (3) (2011) 8016–8026. [33] A.M. Martinez, R. Benavente, The AR Face Database, Tech. Rep., Computer Vision Centre, Autonomous University of Barcelona, June 1998. [34] K. Lee, J. Ho, D. Kriegman, Acquiring linear subspaces for face recognition under variable lighting, IEEE Trans. Pattern Anal. Mach. Intell. 5 (2005) 684–698. [35] T. Sim, S. Baker, M. Bsat, The cmu pose, illumination, and expression database, IEEE Trans. Pattern Anal. Mach. Intell. 25 (12) (2003) 1615–1618. [36] P.J. Phillips, M. Hyeonjoon, S.A. Rizvi, P.J. Rauss, FERET evaluation methodology for face-recognition algorithms, IEEE Trans. Pattern Anal. Mach. Intell. 22 (2000) 1090–1104.

Please cite this article as: J. Li, et al., LEDTD: Local edge direction and texture descriptor for face recognition, Signal Processing-Image Communication (2015), http://dx.doi.org/10.1016/j.image.2015.12.003i

41 43 45 47 49 51 53 55 57 59 61 63 65 67 69 71 73 75 77 79