Rotated Local Binary Pattern (RLBP)

Rotation Invariant Texture Descriptor

Rakesh Mehta and Karen Egiazarian

Tampere University of Technology, Tampere, Finland

Keywords:

Texture Descriptor, Local Binary Pattern (LBP), Rotation Invariance, Local Descriptor.

Abstract:

In this paper we propose two novel rotation invariant local texture descriptors. They are based on Local

Binary Pattern (LBP), which is one of the most effective and frequently used texture descriptor. Although

LBP efﬁciently captures the local structure, it is not rotation invariant. In the proposed methods, a dominant

direction is evaluated in a circular neighbourhood and the descriptor is computed with respect to it. The

weights associated with the neighbouring pixels are circularly shifted with respect to this dominant direction.

Further, in the second descriptor, the uniformity of the patterns is utilized to extract more discriminative

information. The proposed methods are tested for the task of texture classiﬁcation and the performance is

compared with original LBP and its existed extensions.

1 INTRODUCTION

Texture classiﬁcation is an important area of research

in computer vision and pattern recognition. The

texture based descriptors have been used for object,

scene and face classiﬁcation. A number of approaches

can be found in literature based on ﬁlter banks (Salas

and Hille, 1978), co-occurrence statistics (Haralick,

1979), local scale (Ojala et al., 2002) and more re-

cently, on keypoints based setting (Ling and Soatto,

2007) and multi fractal schemes (Xu et al., 2009).

Among them, Local Binary Pattern (LBP) (Ojala

et al., 2002) has gained a popularity due to compu-

tational simplicity and good performance. The origi-

nal LBP operator is invariant to monotonic gray scale

changes as it is computed by taking a difference of the

pixels intensities. However, it is not invariant to im-

age rotations. A number of extensions of LBP have

been proposed to incorporate the rotation invariance

property into it. Ojala et. al. proposed LBPROT

which circularly shifts the binary code until it cor-

responds to one of the preselected rotation invariant

patterns (Ojala et al., 2002). This approach, how-

ever, loses a discriminative information because of

the non-uniformity of the pattern density in the im-

age. To retain the global information Li et. al. pro-

posed a circular shifting of the histogram bins with

respect to the largest bin, since it corresponds to the

most frequent pattern in the image (Li et al., 2012).

Gou et. al. combined the information related to the

128

a b

Figure 1: (a) The neighbourhood for LBP (b) Weights asso-

ciated with the neighbours.

magnitude and central pixel that is complementary to

the sign values (Guo et al., 2010a). They also pro-

posed to estimate the principle orientation and then

align the LBP features with respect to this orientation

(Guo et al., 2010b). Recently Zhao et. al. proposed

LBP Histogram Fourier (LBP-HF) descriptor (Zhao

et al., 2012), that computes the discrete Fourier trans-

form over the histograms to achieve rotation invari-

ance. The Fourier transform, however, completely

ignores the structural arrangement of the histogram,

thereby losing some discriminative information.

In all above mentioned approaches the rotation

invariance is achieved at the expense of losses of

some discriminative features extracted using original

LBP. These losses occur due to the compact map-

ping (Ojala et al., 2002), transformation of histograms

(Zhao et al., 2012) or shifting of the histograms bins

(Li et al., 2012). In this paper we propose a frame-

work which can incorporate the complete structural

497

Mehta R. and Egiazarian K. (2013).

Rotated Local Binary Pattern (RLBP) - Rotation Invariant Texture Descriptor.

In Proceedings of the 2nd International Conference on Pattern Recognition Applications and Methods, pages 497-502

DOI: 10.5220/0004334304970502

 SciTePress

221

170

202

52 127

128

170

221

43 127

128

194

Figure 2: Effect of rotation on LBP operator (a) The image (top) and 90

counter-clockwise rotated image (bottom), (b) In

the rotated images the neighbourhood is rotated counter-clockwise by 90

, (c) Values above threshold are shown in red color,

(d) The weights corresponding to the thresholded neighbours, (e) LBP values.

information extracted by LBP and, at the same time,

achieve a rotation invariance. The main idea behind

the paper is to set a local reference direction in ev-

ery circular neighbourhood and compute the descrip-

tor with respect to it. When the image is rotated, the

local reference direction is also subjected to the rota-

tion by the same degree, hence, the descriptor com-

puted with respect to it remains the same. The choice

of local reference, instead of a global one, is moti-

vated by the fact that the texture is a local cue and is

computed locally by LBP. Usage a single global ref-

erence for a whole image may lead to computation of

LBPs which are misaligned.

This paper starts with a discussion on the reason-

ing behind the inability of LBP to deal with a rotation

changes. Then we introduce the idea of dominant di-

rection, which is used as the reference while comput-

ing the local descriptor. Based on this idea we propose

novel descriptors RLBP and uRLBP. RLBP is com-

puted by circularly shifting the binary code at each

location based on the dominant direction. uRLBP is

an extension of RLBP, utilizing the concept of uni-

form pattern. The proposed descriptors are easy to

compute and with a relatively less number of neigh-

bours they achieve a good performance.

The rest of the paper is organized as follow. In

Section 2 we discuss the LBP operator on rotated

images. In Section 3, we present the proposed ap-

proach for a texture classiﬁcation based on two novel

descriptors RLBP and uRLBP. The experiments are

performed on the standard texture datasets to com-

pare the performance of the proposed methods with

the existed ones. The results are reported in Section 4

and, ﬁnally, the paper is concluded in Section 5.

2 LBP ON ROTATED IMAGES

LBP operator is computed in a local circular region by

taking the difference of the center pixel with respect

to its neighbours. It is deﬁned as

LBP

R,P

P−1

∑

p=0

s(g

− g

) · 2

, (1)

s(g

− g

) =



1 g

≥ g

0 g

< g

. (2)

where g

and g

denote the gray values of the cen-

tral pixel and its neighbour, respectively, p is the in-

dex of the neighbour, R is the radius of the circu-

lar neighbourhood and P is the number of the neigh-

bours. Fig. 1 shows the neighbours for P = 8 and

the weights corresponding to these neighbours. If the

coordinate of the central pixel is (x,y), then the coor-

dinates of uniformly spaced circular neighbourhood

are given as (x + Rcos(2πp/P),y − Rsin(2πp/P)) for

p = 0,1,2,...,P − 1. If the neighbouring coordinate

does not correspond to integer values, then bilinear

interpolation is used for estimation of pixel value.

Further, to extract the most fundamental struc-

ture from the LBP, the idea of uniform pattern is uti-

lized. A local binary pattern is called uniform if bi-

nary code contains at most two transitions from 0 to

1 or vice versa. For example, the patterns 00011100,

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

498

01000000 are uniform as both consist of 2 transitions,

while 00101000 and 00011010 are non-uniform as

they contain 4 transitions. In practice, this is imple-

mented using a lookup table of elements where the

table maps the non-uniform patterns into a single bin

and all others separately. For P neighbours, the num-

ber of uniform LBP patterns is given by P(P − 1)+ 3.

The LBP operator takes the difference of the cen-

tral pixel with the neighbouring pixels and combines

the signs of these differences using unique weights.

The order of the weights is ﬁxed in the circular neigh-

bourhood, i.e. the weight corresponding to g

is al-

ways 1 and so on, as it is shown in Fig. 1. If the im-

age undergoes a rotation, the arrangement of the pixel

around the center undergoes a shift. Since the order of

the weights is ﬁxed, the LBP computed on the rotated

images is unable to deal with the rotation changes.

Thus, even for a simple image rotation the LBP oper-

ator provide very different values. Fig. 2 shows the

effect of rotation on the LBP operator. In this illus-

tration LBP is computed at some location in image

and on a rotated version of the same image. As image

is rotated counter-clockwise by 90

, the neighbours

around each pixel also undergo a counter-clockwise

rotation of the same angle as shown in Fig. 2(b). The

pixels in the circular neighbourhood are thresholded

with respect to the central pixel and those above the

threshold are shown in red color in Fig. 2(c). The

weights corresponding to these pixel values are also

highlighted in red color in Fig. 2(d). The thresholded

pixels and weights in both images correspond to very

different locations and when summed up this results

in distinct output values. It is due to the fact that the

arrangement of the weights in LBP does not depend

on the values of the neighbours. In order to overcome

this problem we propose to adaptively select the ar-

rangement of the weights based on the values of pix-

els in the neighbourhood.

3 TEXTURE CLASSIFICATION

USING RLBP

In this section we present the proposed approach for

the task of texture classiﬁcation. First, we intro-

duce RLBP operator obtained by circularly shifting

the weights of LBP operator. Further, the intrin-

sic structure of the patterns is utilized by incorporat-

ing the principle of uniform patterns to generate uni-

form RLBP (uRLBP). Finally, two classiﬁers (nearest

neighbour (NN) and support vector machine (SVM))

are discussed along with their choice of parameters.

3.1 Rotated Local Binary Pattern

(RLBP)

LBP only considers the signs of the differences to

compute the ﬁnal descriptor. The information related

to the magnitude of the differences is completely ig-

nored. The magnitude provides a complimentary in-

formation that has been utilized (Guo et al., 2010a)

to increase the discriminative power of the operator.

Especially in the neighbourhood with strong edges

the magnitude of the differences can provide an im-

portant information. Here we utilize the magnitude

of the difference to ﬁnd the dominant direction in a

neighbourhood. The dominant direction is deﬁned as

the index in the circular neighbourhood for which the

difference is maximum. As an image undergoes a ro-

tation the dominant direction in a neighbourhood also

undergoes the rotation by the same angle. In the pro-

posed descriptor the dominant direction is set as the

reference and the weights for the neighbourhood are

arranged with respect to it.

In order to make the LBP invariant to rotation we

circularly shift the weights according to the dominant

direction. The dominant direction (D) in a neighbour-

hood is the index of neighbour whose difference to

the central pixel is maximum; it is deﬁned as

D = argmax

p∈(0,1...P−1)

− g

|. (3)

The rotation of neighbourhood with respect to its cen-

ter shifts the direction D by the same angle. For ex-

ample, if we consider a neighbourhood [23 25 28;

167 35 31; 56 67 72], after the thresholding its bi-

nary code is [01111000], the index of D is 4 (shown

in bold) corresponding to the pixel value 167 . If the

image is rotated by 45

counter-clockwise, then the

neighbourhood also rotates by the same angle and the

binary code shifts to [00111100]. D index still cor-

responds to the same value (167) but it shifts by one

step. However, the circular arrangement of the pix-

els in the neighbourhood remains the same with re-

spect to the direction D. Therefore the lowest weight

is associated with the index corresponding to D and is

subsequently increased in the counter-clockwise di-

rection.

Since the dominant direction is taken as the ref-

erence in the circular neighbourhood, the weights are

assigned with respect to it. Thus the RLBP operator

is deﬁned as

RLBP

R,P

P−1

∑

p=0

s(g

− g

) · 2

mod(p−D,P)

, (4)

where mod indicates the modulus operation. In the

above deﬁnition the weight term 2

mod(p−D,P)

depends

RotatedLocalBinaryPattern(RLBP)-RotationInvariantTextureDescriptor

499

221

170

202

52 127

128

170

202

221

43 127

128

Figure 3: Effect of rotation on RLBP operator (a) The image (top) and 90

counter-clockwise rotated image (bottom), (b)

Yellow color pixel indicate the dominant direction, (c) Values above threshold are shown in red color, (d) The weights are

circularly shifted with respect to dominant direction, (e) RLBP values.

on D. The weights are circularly shifted with respect

to the dominant direction. The shift results in a ro-

tation invariance, as the weights now depend on the

neighbourhood and not on a preselected arrangement.

Fig. 3 shows the effect of rotation on the RLBP fea-

tures. For clear understanding we use the same sce-

nario as was used in Fig. 2 for LBP. As before, the red

color indicate the pixels above the threshold, yellow

color indicate the pixel corresponding to the dominant

direction D. The bit corresponding to index D always

takes the lowest weight of 1 and other weights are cir-

cularly shifted with respect to it. In Fig. 3(d) it can

be seen that the weight corresponding to the dominant

position is the same for original and rotated images,

although these pixels are at different locations. Thus

the RLBP values obtained for two different rotated

neighbourhood are similar in this case.

3.2 Uniform Rotated Local Binary

Pattern

To utilize the intrinsic structure of the binary pat-

terns the idea of uniform patterns is applied on the

LBP. The uniform LBP (uLBP) achieves better per-

formance compared to LBP due to the statistical prop-

erties of these patterns. Experiments carried out on

large image datasets showed that up to 90% of the

total patterns are uniform while the remaining small

percentage are non uniform. This small portion of

non-uniform patterns are distributed in large number

of histogram bins which cannot be estimated reliably.

The accumulation of large number of uniform pat-

terns into relatively small number of histogram bins

provides the discriminative information, which is uti-

lized by uLBP. Thus, if the distribution of the patterns

obtained by the operator is preserved, the uniformity

of the pattern can be utilized to enhance the discrimi-

native power.

Similar to the LBP operator, the RLBP operator

computes the binary patterns based on pixel neigh-

bourhood. The ﬁnal values obtained by these oper-

ators are different because the weights are circularly

shifted in RLBP operator. Since the binary patterns

are same for both operators, the distribution of the

patterns is similar. To capture an additional discrimi-

native information, we apply the idea of the uniform

patterns to RLBP operator. The uniform RLBP opera-

tors is given as U(RLBP

R,P

), where U() is the lookup

table that deﬁnes a mapping from a binary pattern to

a uniform pattern. For a neighbourhood of P pix-

els, the lookup table U () consist of 2

elements and

P(P − 1) + 3 different output values corresponding to

the uniform patterns.

3.3 Classiﬁers

The performance of the descriptor is tested by two dif-

ferent classiﬁers: Nearest Neighbour (NN) and Sup-

port Vector Machine (SVM). The nearest neighbour

classiﬁer is used with a number of different distance

metrics such as chi-square, log-likehood, cosine dis-

tance, etc. In this study we use the histogram inter-

section which is given as

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

500

d(H,K) = 1 −

∑

min(h

)

∑

(5)

where H is the histogram of the test sample, K is the

histogram of the training sample, i represent the bin

number and h

, k

are the values of the the i

bin in

the histogram H and K respectively. The test sample

is assigned to the class of the training sample which

minimizes this distance.

Nearest neighbour is a simple classifer which can

deal with easy rotational variations. However it can-

not effectively model the texture with more difﬁcult

viewpoint and scale variations. Thus, to deal with

these difﬁcult conditions, SVM classiﬁer is utilized.

The SVM classiﬁer is tested with different kernels and

the linear kernels provided the best numerical results.

4 EXPERIMENTS

The proposed method is tested on three different tex-

ture datasets: Outex-10, Outex-12 and UIUC tex-

ture datasets. Outex-10 and Outex-12 have only ro-

tational variation while the UIUC is much more dif-

ﬁcult dataset with rotation, scale and viewpoints vari-

ations. The performance of the proposed algorithm

is compared with LBP, uniform LBP (uLBP), rota-

tion invariant LBP (riLBP), uniform rotation invari-

ant LBP (uriLBP) and LBP-HF.Three of these meth-

ods ((riLBP),(uriLBP) and LBP-HF) are rotation in-

variant version of LBP. To keep the complexity low

we set the number of neighbours to 8 in all the tests.

4.1 Outex-12

Outex-12 dataset consists of 9120 images represent-

ing 24 different textures under different lighting and

rotations. In our experiments, 20 images from each

class are used for training and the rest are used for

testing. Thus, the testing set consists of 8640 images

and training set consists of 480 images. The results of

comparison are shown in Table 1. It can be observed

that the methods proposed speciﬁcally to deal with the

rotation changes, such as riLBP, LBP-HF and the pro-

posed descriptor, achieve much better accuracy than

the original LBP. LBP-HF preforms better than ear-

lier proposed descriptors riLBP and uriLBP. However,

the proposed descriptor outperforms this state-of-the-

art rotation invariant LBP-HF. The radius parameter

is increased from 1 to 3 for LBP-HF and RLBP. It is

interesting to observe that for both descriptors the per-

formance improves with an increase in the radius. At

smaller radius the operator captures pixel related in-

formation while at the larger radius the region based

Table 1: Test results for Outex-12 dataset.

Method Accuracy Dimensionality

LBP 54.8 256

uLBP 54.54 59

riLBP 66.40 36

uriLBP 62.23 10

LBP − HF

1,8

75.48 38

LBP − HF

2,8

80.30 38

LBP − HF

3,8

82.97 38

RLBP

1,8

72.35 256

RLBP

2,8

80.88 256

RLBP

3,8

85.68 256

uRLBP

1,8

74.86 59

uRLBP

2,8

83.19 59

uRLBP

3,8

88.01 59

information is captured. For radius 3 the operator cap-

tures the directional information from the circular re-

gions of diameter 7. As the radius is increased further

the accuracy decreases, which implies that for large

region (radius > 3) binary patterns cannot effectively

capture the structure.

4.2 Outex-10

The Outex-10 dataset consists of 4320 images

belonging to 24 different textures classes. The

images are rotated at nine different angles

,10

,15

,30

,45

,60

,75

,90

) and the

illumination is kept constant. In our experiments,

20 images from each class are used for training

and the rest 3840 images are used for testing. The

results are shown in Table 2. LBP and uLBP perform

poorly because they are not designed to deal with

any rotational variations. Their rotation invariant

versions, riLBP and uriLBP, provide much better

results. The best accuracy is again achieved by the

uRLBP with radius 3. The proposed method again

achieves higher accuracy than the recently proposed

LBP-HF which can be attributed to the fact that the

structural information is retained in the RLBP while

in the LBP-HF rotation invariance is achieved at

the loss of the structural information of the binary

patterns.

4.3 UIUC Dataset

In order to test descriptor against more difﬁcult con-

ditions we have used the UIUC texture dataset. This

dataset consist of 1,000 uncalibrated unregistered im-

ages: 40 samples each of 25 different textures. Sig-

niﬁcant rotation, viewpoint changes and scale differ-

ences are present within each class and illumination

RotatedLocalBinaryPattern(RLBP)-RotationInvariantTextureDescriptor

501

Table 2: Test results for Outex-10 dataset.

Method Accuracy Dimensionality

LBP 52.66 256

uLBP 54.19 59

riLBP 83.15 36

uriLBP 84.01 10

LBP − HF

1,8

78.46 38

LBP − HF

2,8

83.46 38

LBP − HF

3,8

87.81 38

RLBP

1,8

86.41 256

RLBP

2,8

89.82 256

RLBP

3,8

90.39 256

uRLBP

1,8

88.07 59

uRLBP

2,8

92.94 59

uRLBP

3,8

95.96 59

conditions are uncontrolled. In our experiments we

use 10 images from each class for training and rest

of the images for testing. The results of the compari-

son are shown in Table 3. Although the descriptor is

not speciﬁcally designed to deal with viewpoint and

scale changes, it achieves best performance among

the rotation invariant LBP versions. It can be ob-

served that the performance of other rotation invariant

LBP (uLBP, riLBP and LBP-HF) decreases consid-

erable on this dataset. Proposed descriptor achieves

highest accuracy because it retains the structural in-

formation captured by the original LBP in addition

the rotation invariance helps to deal with the view-

point changes. Also it is interesting to note the role

of uniform pattern in this dataset; for all three de-

scriptors LBP, riLBP and RLBP where uniform pat-

terns are utilized, the accuracy drops by few percent

for their uniform counterpart. In the difﬁcult rotation,

scale and viewpoint variations the small number of

uniform patterns are not sufﬁcient to model the neigh-

bourhood of the pixels.

5 CONCLUSIONS

In this paper we have presented rotation invariant lo-

cal descriptors for texture classiﬁcation. The opera-

tors utilize information from the local neighbourhood

to achieve a rotation invariance. Under any rotation

transformation the local circular neighbourhood also

undergoes the rotation around its center. However, if

the descriptor is computed by using the local dom-

inant direction as the reference, then it becomes in-

variant to a rotation. Based on this idea we proposed

two rotation invariant descriptors RLBP and uRLBP.

The experiments performed on the standard texture

datasets show that the proposed method performs bet-

Table 3: Test results for UIUC Texture dataset.

Method Accuracy Dimensionality

LBP 50.40 256

uLBP 49.2 59

riLBP 58.00 36

uriLBP 56.13 10

LBP − HF

1,8

59.73 38

LBP − HF

2,8

70.00 38

LBP − HF

3,8

66.40 38

RLBP

1,8

67.06 256

RLBP

2,8

73.60 256

RLBP

3,8

77.47 256

uRLBP

1,8

67.06 59

uRLBP

2,8

72.53 59

uRLBP

3,8

74.53 59

ter than a number of state-of-the-art LBP based rota-

tion invariant descriptors.

REFERENCES

Guo, Z., Zhang, L., and Zhang, D. (2010a). A completed

modeling of local binary pattern operator for texture

classiﬁcation. Image Processing, IEEE Transactions

on, 19(6):1657–1663.

Guo, Z., Zhang, L., and Zhang, D. (2010b). Rota-

tion invariant texture classiﬁcation using lbp variance

(lbpv) with global matching. Pattern Recognition,

43(3):706–719.

Haralick, R. (1979). Statistical and structural approaches to

texture. Proceedings of the IEEE, 67(5):786–804.

Li, Z., Liu, G., Yang, Y., and You, J. (2012). Scale-

and rotation-invariant local binary pattern using

scale-adaptive texton and subuniform-based circular

shift. Image Processing, IEEE Transactions on,

21(4):2130–2140.

Ling, H. and Soatto, S. (2007). Proximity distribution ker-

nels for geometric context in category recognition. In

IEEE ICCV., pages 1–8.

Ojala, T., Pietikainen, M., and Maenpaa, T. (2002). Mul-

tiresolution gray-scale and rotation invariant texture

classiﬁcation with local binary patterns. IEEE TPAMI,

24(7):971 –987.

Salas, S. and Hille, E. (1978). Calculus: One and Several

Variable. John Wiley and Sons, New York.

Xu, Y., Ji, H., and Ferm

uller, C. (2009). Viewpoint invari-

ant texture description using fractal analysis. IJCV,

83(1):85–100.

Zhao, G., Ahonen, T., Matas, J., and Pietikainen, M. (2012).

Rotation-invariant image and video description with

local binary pattern features. Image Processing, IEEE

Transactions on, 21(4):1465 –1477.

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

502