An Effective Texture Spectrum Descriptor
2009 Fifth International Conference on Information Assurance and Security
An Effective Texture Spectrum Descriptor Wu Xiaosheng1
Sun Junding1,2
1
2
School of Computer Science and technology Henan Polytechnic University Jiaozuo 454003, China [email protected]
Provincial Key Laboratory of Computer Information Processing Technology Suzhou University Suzhou, 215006, China [email protected]
Abstract—The center-symmetric local binary pattern (CS-LBP) is an effective extension to local binary pattern (LBP) operator. However, it discards some important texture information because of the ignorance of the center pixel and is hard to choose a threshold for recognizing the flat area. A novel improved CS-LBP operator, named ICS-LBP, is proposed in this paper. The new operator classifies the local pattern based on the relativity of the center pixel and the center-symmetric pixels instead of the gray value differences between the centersymmetric pixels as CS-LBP, which can fully extract the texture information discarded by CS-LBP descriptor. Comparisons are given among the three methods and the experimental results show the performance improvement of the new descriptor. Keywords-image retrieval; texture spetrum; LBP; CS-LBP; ICS-LBP
I.
INTRODUCTION
Image texture analysis is an important fundamental task in computer vision and pattern recognition. More currently, the local binary pattern (LBP) has received considerable attentions in many applications, such as face recognition [1], image retrieval [2] and motion analysis [3], for its efficiency and simple theory and computation. The LBP texture analysis operator is defined as a grayscale invariant texture measure, derived from a general definition of texture in a local neighborhood. For each pixel in an image, a binary code is produced by thresholding its value with the value of the center pixel. A histogram is created to collect up the occurrences of different binary patterns. The basic version of the LBP operator considers only the eight neighbors of a pixel [4], but the definition has been extended to include all circular neighborhoods with any number of pixels [5]. In [6], Heikkilä introduced the CS-LBP operator for region description. However, the center pixel was ignored in this method. Therefore, it can not describe the gradient of the neighborhood efficiently. In this paper, we introduce an improved CS-LBP, named ICS-LBP, which considers the relativity of the center pixel and the pairs of the opposite pixels in a circle neighborhood, not on the gray-level differences between pairs of opposite pixels as the CS-LBP descriptor. Comparisons are given among LBP, CS-LBP and ICS-LBP on three texture
978-0-7695-3744-3/09 $25.00 © 2009 IEEE DOI 10.1109/IAS.2009.126
databases with different evaluation criterions in the experiments. The rest of the paper is organized as follows. In Section 2, we first briefly describe the LBP and CS-LBP methods, and then introduce the proposed descriptor in details. The experimental results and analyses are described in Section 3. Finally, we conclude the study in Section 4. II.
TEXTURE SPECTRUM
Before presenting ICS-LBP descriptor in details, we give a brief review of the LBP and CS-LBP firstly as the basis of our work. A. LBP & CS-LBP LBP operated with eight neighboring pixels using the center as a threshold. If the gray-level of the neighboring pixels is higher or equal, the value is set to one, otherwise to zero. The final LBP code was then produced by multiplying the thresholded values by weights given by powers of two and adding the results in a way described by figure 1. By definition, LBP is invariant to any monotonic transformation of the gray scale and it is quick to compute. The operators with different parameters can be combined to obtain a multiscale description of textures, see figure 2 for an illustration [7]. In figure 2, three neighborhoods with a varying number of samples (P) and different neighborhood radii (R) are shown, which can be denoted by LBPP , R .
Figure 1. The computation of original LBP
Figure 2. Circularly symmetric neighbor sets
361
However, the LBP descriptor produces rather long histograms (256) and is therefore difficult to use as a region descriptor. Heikkilä proposed the CS-LBP descriptor [6] for region description. Different from LBP, they did not compare the gray-level of each pixel with gray-level of the center pixel, but the gray-level difference of the pairs of opposite pixels. At the same time, a threshold was set to obtain the robustness on flat image regions.
CS _ LBPP , R ,T ( x, y ) =
⎧⎪1, pi − pi + ( P 2) > T sCS − LBP ( pi , pi + ( P 2) ) = ⎨ ⎪⎩0, otherwise
=
i =0
sCS − LBP ( pi , pi + ( P 2) ) × 2i
(5)
∑
sICS − LBP ( pi , pc , pi + ( P 2) ) × 2i
(6)
where ( x, y ) denotes the coordinates of a pixel. An illustration for LBP, CS-LBP and ICS-LBP is given in figure 3 for 8 neighbors. It is clear that LBP produces 256 (28) different binary patterns, whereas CS-LBP and ICS-LBP produce only 16 (24) different binary pattern for 8 neighbors.
Figure 3. The LBP, CS-LBP and ICS-LBP features for a neighborhood of 8 pixels.
III.
EXPERIMENTAL EVALUATION
In this section, three data sets were used as the test beds. We set P = 8 and. For the CS-LBP descriptor, the threshold T is set 0.01, which has been proved efficiently in the reference [7]. L1 distance is chosen as the measurement. The first texture data set involves 16 class textures from the Brodatz album, which are shown in figure 4. Each class of the texture images is presented at 10 different rotation angles (0o, 20o, 30o, 45o, 60o, 70o, 90o, 120o, 135o and 150o), totally 160 texture images which are downloaded from http://www.ee.oulu.fi/mvg/page/. The second texture volume consists of 12 class textures from the Brodatz texture images, which are shown in figure 5. Each texture is digitized at seven different rotation angles (0 o, 30 o, 60 o, 90 o, 120 o, 150 o and 200 o), totally 84 texture images. The data set is downloaded from http://sipi.usc.edu/database/ . The third database is made up of the first, the second database and the other texture images downloaded from the Internet, totally 550 texture images. All the images in the three databases are fixed to 64×64 pixels in the experiments. For the first and the second databases, PN - the precision of the first N retrieved images and RN -the recall of the first N retrieved images are chosen as the evaluation criterion [8][9], which are defined as (Liu, et al., 2001, Ru, et al., 2003)
(1)
(2)
⎧⎪1, (pi ≥ pc & pc ≥ pi + ( P 2) ) | (pi < pc & pc < pi + ( P 2) ) (3) =⎨ ⎪⎩0, otherwise
where pi , pi + ( P 2) and pc correspond to the gray-level of center-symmetric pairs of pixels and the center pixel on a circle of radius R , T is the threshold for the CS-LBP descriptor. For the meaning of Eq. 1 and Eq. 2, refer to [6]. By the function above, the binary pattern of LBP, CS-LBP and ICS-LBP are calculated as P −1
( P 2) −1 i =0
sICS − LBP ( pi , pc , pi + ( P 2) )
LBPP , R ( x, y ) = ∑ sLBP ( pi , pc ) × 2i
∑
ICS _ LBPP , R ( x, y )
B. Improved CS-LBP The template is used to format your paper and style the text. All margins, column widths, line spaces, and text fonts are prescribed; please do not alter them. You may note peculiarities. For example, the head margin in this template measures proportionately more than is customary. This measurement and others are deliberate, using specifications that anticipate your paper as one part of the entire proceedings, and not as an independent document. Please do not revise any of the current designations. Though CS-LBP is a more efficient region descriptor than LBP, the gradient information is not considered entirely because of the ignorance of the center pixel. It is also hard to choose an adaptable threshold. To address the problems, we proposed ICS-LBP descriptor. In the new method, we consider the relation of the center pixel and the centersymmetric pairs of pixels instead of the gray-level difference between the center-symmetric pairs as CS-LBP. The scheme functions of LBP, CS-LBP and ICS-LBP are given together as ⎧1, pi ≥ pc sLBP ( pi , pc ) = ⎨ ⎩0, otherwise
( P 2) −1
(4)
i =0
Pi , N (q ) = ∑ j =1 N
362
φ ( I j , Ri ) N
(9)
Ri , N (q ) = ∑ j =1 N
φ ( I j , Ri ) Ri
as query images, the average recall-precision graph averaged over 84 queries for the 3 descriptors are shown in by figure 9. The same results can be concluded as figure 8
(10)
IV.
where Ri denotes the i-class texture which contains Ri
CONCLUSION
In this paper, we introduced the ICS-LBP descriptor for texture image retrieval, which combines the strengths of the SIFT descriptor and the LBP operator. It was evaluated against the LBP and CS-LBP descriptors on three texture databases with different evaluation criterions. The experimental results show that the ICS-LBP descriptor is better for most of the test cases than the other two descriptors. Future work includes applying the proposed descriptor to different computer vision problems such as object recognition and other applications.
images, and q is a query image, and I 1 , I 2 , , I j , , I N are the first N retrieved images and ⎧⎪0, if I j ∉ Ri . φ ( I j , Ri ) = ⎨ ⎪⎩1, if I j ∈ Ri
0.7 CS-LBP LBP ICS-LBP
0.6
precision
0.5 0.4 0.3 0.2 0.1 0 0
Figure 4. 16 Brodatz textures in the first database
2
4
(a) Average
6
8
10
12
class
14
16
18
PN of the 3 methods for N = 15
0.7 CS-LBP LBP ICS-LBP
0.6
precisio
0.5 0.4 0.3 0.2 0.1
Figure 5. 12 Brodatz textures in the second database
0
In the experiments, all the images in each database are selected as query images. The average PN and RN of the 3 descriptors for each texture class in the two databases are illustrated in figure 6 and figure 8 respectively. The results in figure 6 show that the ICS-LBP descriptor outperforms CSLBP descriptor except 8- and 16-class texture images and it is better than LBP descriptor except 3- and 8-class. For most of the test cases, LBP descriptor is able to outperform CSLBP descriptor. The same conclusion can be proved by the results tested on the second texture database in figure 7. For the third texture database, the retrieval efficiency was measured in terms of the recall and precision. Firstly, we use the texture images belonging to the first database as queries, the average recall-precision graph averaged over 160 queries for the 3 descriptors are shown in figure 8. It also denotes that the ICS-LBP descriptor is more preferable than CS-LBP and LBP descriptor in most of the test cases. Secondly, the texture images belonging to the second database are selected
1
2
3
(b) Average
4
5
6
7
class
8
9
10
11
12
RN of the 3 methods for N = 15
Figure 6. Comparison on the first database 1
0.8
CS-LBP LBP ICS-LBP
recall
0.6
0.4
0.2
0 0
2
4
6
8
10
class
12
14
16
(a) Average PN of the 3 methods for N = 10
363
18
1 CS-LBP LBP ICS-LBP
0.8
ACKNOWLEDGMENT This work was supported by NSF of education bureau of Henan province 2008B520012, 2008B520015, 2009B520013, the Internal Cooperation Science Foundation of Henan province 084300510065, the Dr. and backbone teacher grant of Henan Polytechnic University, Provincial Key Laboratory of Computer Information Processing Technology of Suzhou University KJS0715.
recall
0.6
0.4
0.2
0
1
2
3
4
5
6
class
7
8
9
10
11
12
(b) Average RN of the 3 methods for N = 10
REFERENCES
Figure 7. Comparison on the second database [1] 1 CS-LBP LBP ICS-LBP
precision
0.8
[2]
0.6
[3]
0.4
0.2
0 0.1
[4] 0.2
0.3
0.4
0.5
0.6
recall
0.7
0.8
0.9
1
[5]
Figure 8. Average recall-precision graph averaged over 160 query images 1
[6]
CS-LBP LBP ICS-LBP
0.8
precision
[7] 0.6
[8]
0.4
0.2
[9] 0 0.1
0.2
0.3
0.4
0.5
0.6
recall
0.7
0.8
0.9
1
Figure 9. Average recall-precision graph averaged over 91 query images
364
Y.K. Park and J.K. Kim, “Fast adaptive smoothing based on LBP for robust face recognition”, IEE Electronics Letters, 2007, 43 (24):13501351. K. Mikolajczyk and C. Schmid, “Indexing based on scale invariant interest points”, in: 8th IEEE International Conference on Computer Vision, 2001, pp. 525-531. M. Heikkilä and M. Pietikäinen, “A texture-based method for modeling the background and detecting moving objects”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2006, 28(4):657-662. T. Ojala, M. Pietikäinen, and D. Hardwood, “A comparative study of texture measures with classification based on feature distribution”, Pattern Recognition, 1996, 29:51-59. T. Ojala, M. Pietikäinen, and T. Mäenpää, “Multiresolution grayscale and rotation invariant texture classification with Local Binary Patterns”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2002, 24(7):971- 987. M. Heikkilä, M. Pietikäinen, and C. Schmid, “Description of interest regions with local binary patterns”, Pattern Recognition, 2009,42(3):425-436. T. Mäenpää and M. Pietikäinen, “Multi-scale binary patterns for texture analysis”, Lecture Notes in Computer Science 2749, pp. 885892. W.Y. Liu, Z. Su, S. Li, et al, “A performance evaluation protocol for content-based image retrieval algorithms/systems”, In CD-ROM of IEEE CVPR Workshop on Empirical Evaluation Methods in Computer Vision, Hawaii, USA. L.Y. Ru, X. Peng, Z. Su, et al, “Feature performance evaluation in content-based image retrieval”, Journal of computer research and development, 2003, 40(11):1566-1560. (In Chinese)
An Effective Texture Spectrum Descriptor Wu Xiaosheng1
Sun Junding1,2
1
2
School of Computer Science and technology Henan Polytechnic University Jiaozuo 454003, China [email protected]
Provincial Key Laboratory of Computer Information Processing Technology Suzhou University Suzhou, 215006, China [email protected]
Abstract—The center-symmetric local binary pattern (CS-LBP) is an effective extension to local binary pattern (LBP) operator. However, it discards some important texture information because of the ignorance of the center pixel and is hard to choose a threshold for recognizing the flat area. A novel improved CS-LBP operator, named ICS-LBP, is proposed in this paper. The new operator classifies the local pattern based on the relativity of the center pixel and the center-symmetric pixels instead of the gray value differences between the centersymmetric pixels as CS-LBP, which can fully extract the texture information discarded by CS-LBP descriptor. Comparisons are given among the three methods and the experimental results show the performance improvement of the new descriptor. Keywords-image retrieval; texture spetrum; LBP; CS-LBP; ICS-LBP
I.
INTRODUCTION
Image texture analysis is an important fundamental task in computer vision and pattern recognition. More currently, the local binary pattern (LBP) has received considerable attentions in many applications, such as face recognition [1], image retrieval [2] and motion analysis [3], for its efficiency and simple theory and computation. The LBP texture analysis operator is defined as a grayscale invariant texture measure, derived from a general definition of texture in a local neighborhood. For each pixel in an image, a binary code is produced by thresholding its value with the value of the center pixel. A histogram is created to collect up the occurrences of different binary patterns. The basic version of the LBP operator considers only the eight neighbors of a pixel [4], but the definition has been extended to include all circular neighborhoods with any number of pixels [5]. In [6], Heikkilä introduced the CS-LBP operator for region description. However, the center pixel was ignored in this method. Therefore, it can not describe the gradient of the neighborhood efficiently. In this paper, we introduce an improved CS-LBP, named ICS-LBP, which considers the relativity of the center pixel and the pairs of the opposite pixels in a circle neighborhood, not on the gray-level differences between pairs of opposite pixels as the CS-LBP descriptor. Comparisons are given among LBP, CS-LBP and ICS-LBP on three texture
978-0-7695-3744-3/09 $25.00 © 2009 IEEE DOI 10.1109/IAS.2009.126
databases with different evaluation criterions in the experiments. The rest of the paper is organized as follows. In Section 2, we first briefly describe the LBP and CS-LBP methods, and then introduce the proposed descriptor in details. The experimental results and analyses are described in Section 3. Finally, we conclude the study in Section 4. II.
TEXTURE SPECTRUM
Before presenting ICS-LBP descriptor in details, we give a brief review of the LBP and CS-LBP firstly as the basis of our work. A. LBP & CS-LBP LBP operated with eight neighboring pixels using the center as a threshold. If the gray-level of the neighboring pixels is higher or equal, the value is set to one, otherwise to zero. The final LBP code was then produced by multiplying the thresholded values by weights given by powers of two and adding the results in a way described by figure 1. By definition, LBP is invariant to any monotonic transformation of the gray scale and it is quick to compute. The operators with different parameters can be combined to obtain a multiscale description of textures, see figure 2 for an illustration [7]. In figure 2, three neighborhoods with a varying number of samples (P) and different neighborhood radii (R) are shown, which can be denoted by LBPP , R .
Figure 1. The computation of original LBP
Figure 2. Circularly symmetric neighbor sets
361
However, the LBP descriptor produces rather long histograms (256) and is therefore difficult to use as a region descriptor. Heikkilä proposed the CS-LBP descriptor [6] for region description. Different from LBP, they did not compare the gray-level of each pixel with gray-level of the center pixel, but the gray-level difference of the pairs of opposite pixels. At the same time, a threshold was set to obtain the robustness on flat image regions.
CS _ LBPP , R ,T ( x, y ) =
⎧⎪1, pi − pi + ( P 2) > T sCS − LBP ( pi , pi + ( P 2) ) = ⎨ ⎪⎩0, otherwise
=
i =0
sCS − LBP ( pi , pi + ( P 2) ) × 2i
(5)
∑
sICS − LBP ( pi , pc , pi + ( P 2) ) × 2i
(6)
where ( x, y ) denotes the coordinates of a pixel. An illustration for LBP, CS-LBP and ICS-LBP is given in figure 3 for 8 neighbors. It is clear that LBP produces 256 (28) different binary patterns, whereas CS-LBP and ICS-LBP produce only 16 (24) different binary pattern for 8 neighbors.
Figure 3. The LBP, CS-LBP and ICS-LBP features for a neighborhood of 8 pixels.
III.
EXPERIMENTAL EVALUATION
In this section, three data sets were used as the test beds. We set P = 8 and. For the CS-LBP descriptor, the threshold T is set 0.01, which has been proved efficiently in the reference [7]. L1 distance is chosen as the measurement. The first texture data set involves 16 class textures from the Brodatz album, which are shown in figure 4. Each class of the texture images is presented at 10 different rotation angles (0o, 20o, 30o, 45o, 60o, 70o, 90o, 120o, 135o and 150o), totally 160 texture images which are downloaded from http://www.ee.oulu.fi/mvg/page/. The second texture volume consists of 12 class textures from the Brodatz texture images, which are shown in figure 5. Each texture is digitized at seven different rotation angles (0 o, 30 o, 60 o, 90 o, 120 o, 150 o and 200 o), totally 84 texture images. The data set is downloaded from http://sipi.usc.edu/database/ . The third database is made up of the first, the second database and the other texture images downloaded from the Internet, totally 550 texture images. All the images in the three databases are fixed to 64×64 pixels in the experiments. For the first and the second databases, PN - the precision of the first N retrieved images and RN -the recall of the first N retrieved images are chosen as the evaluation criterion [8][9], which are defined as (Liu, et al., 2001, Ru, et al., 2003)
(1)
(2)
⎧⎪1, (pi ≥ pc & pc ≥ pi + ( P 2) ) | (pi < pc & pc < pi + ( P 2) ) (3) =⎨ ⎪⎩0, otherwise
where pi , pi + ( P 2) and pc correspond to the gray-level of center-symmetric pairs of pixels and the center pixel on a circle of radius R , T is the threshold for the CS-LBP descriptor. For the meaning of Eq. 1 and Eq. 2, refer to [6]. By the function above, the binary pattern of LBP, CS-LBP and ICS-LBP are calculated as P −1
( P 2) −1 i =0
sICS − LBP ( pi , pc , pi + ( P 2) )
LBPP , R ( x, y ) = ∑ sLBP ( pi , pc ) × 2i
∑
ICS _ LBPP , R ( x, y )
B. Improved CS-LBP The template is used to format your paper and style the text. All margins, column widths, line spaces, and text fonts are prescribed; please do not alter them. You may note peculiarities. For example, the head margin in this template measures proportionately more than is customary. This measurement and others are deliberate, using specifications that anticipate your paper as one part of the entire proceedings, and not as an independent document. Please do not revise any of the current designations. Though CS-LBP is a more efficient region descriptor than LBP, the gradient information is not considered entirely because of the ignorance of the center pixel. It is also hard to choose an adaptable threshold. To address the problems, we proposed ICS-LBP descriptor. In the new method, we consider the relation of the center pixel and the centersymmetric pairs of pixels instead of the gray-level difference between the center-symmetric pairs as CS-LBP. The scheme functions of LBP, CS-LBP and ICS-LBP are given together as ⎧1, pi ≥ pc sLBP ( pi , pc ) = ⎨ ⎩0, otherwise
( P 2) −1
(4)
i =0
Pi , N (q ) = ∑ j =1 N
362
φ ( I j , Ri ) N
(9)
Ri , N (q ) = ∑ j =1 N
φ ( I j , Ri ) Ri
as query images, the average recall-precision graph averaged over 84 queries for the 3 descriptors are shown in by figure 9. The same results can be concluded as figure 8
(10)
IV.
where Ri denotes the i-class texture which contains Ri
CONCLUSION
In this paper, we introduced the ICS-LBP descriptor for texture image retrieval, which combines the strengths of the SIFT descriptor and the LBP operator. It was evaluated against the LBP and CS-LBP descriptors on three texture databases with different evaluation criterions. The experimental results show that the ICS-LBP descriptor is better for most of the test cases than the other two descriptors. Future work includes applying the proposed descriptor to different computer vision problems such as object recognition and other applications.
images, and q is a query image, and I 1 , I 2 , , I j , , I N are the first N retrieved images and ⎧⎪0, if I j ∉ Ri . φ ( I j , Ri ) = ⎨ ⎪⎩1, if I j ∈ Ri
0.7 CS-LBP LBP ICS-LBP
0.6
precision
0.5 0.4 0.3 0.2 0.1 0 0
Figure 4. 16 Brodatz textures in the first database
2
4
(a) Average
6
8
10
12
class
14
16
18
PN of the 3 methods for N = 15
0.7 CS-LBP LBP ICS-LBP
0.6
precisio
0.5 0.4 0.3 0.2 0.1
Figure 5. 12 Brodatz textures in the second database
0
In the experiments, all the images in each database are selected as query images. The average PN and RN of the 3 descriptors for each texture class in the two databases are illustrated in figure 6 and figure 8 respectively. The results in figure 6 show that the ICS-LBP descriptor outperforms CSLBP descriptor except 8- and 16-class texture images and it is better than LBP descriptor except 3- and 8-class. For most of the test cases, LBP descriptor is able to outperform CSLBP descriptor. The same conclusion can be proved by the results tested on the second texture database in figure 7. For the third texture database, the retrieval efficiency was measured in terms of the recall and precision. Firstly, we use the texture images belonging to the first database as queries, the average recall-precision graph averaged over 160 queries for the 3 descriptors are shown in figure 8. It also denotes that the ICS-LBP descriptor is more preferable than CS-LBP and LBP descriptor in most of the test cases. Secondly, the texture images belonging to the second database are selected
1
2
3
(b) Average
4
5
6
7
class
8
9
10
11
12
RN of the 3 methods for N = 15
Figure 6. Comparison on the first database 1
0.8
CS-LBP LBP ICS-LBP
recall
0.6
0.4
0.2
0 0
2
4
6
8
10
class
12
14
16
(a) Average PN of the 3 methods for N = 10
363
18
1 CS-LBP LBP ICS-LBP
0.8
ACKNOWLEDGMENT This work was supported by NSF of education bureau of Henan province 2008B520012, 2008B520015, 2009B520013, the Internal Cooperation Science Foundation of Henan province 084300510065, the Dr. and backbone teacher grant of Henan Polytechnic University, Provincial Key Laboratory of Computer Information Processing Technology of Suzhou University KJS0715.
recall
0.6
0.4
0.2
0
1
2
3
4
5
6
class
7
8
9
10
11
12
(b) Average RN of the 3 methods for N = 10
REFERENCES
Figure 7. Comparison on the second database [1] 1 CS-LBP LBP ICS-LBP
precision
0.8
[2]
0.6
[3]
0.4
0.2
0 0.1
[4] 0.2
0.3
0.4
0.5
0.6
recall
0.7
0.8
0.9
1
[5]
Figure 8. Average recall-precision graph averaged over 160 query images 1
[6]
CS-LBP LBP ICS-LBP
0.8
precision
[7] 0.6
[8]
0.4
0.2
[9] 0 0.1
0.2
0.3
0.4
0.5
0.6
recall
0.7
0.8
0.9
1
Figure 9. Average recall-precision graph averaged over 91 query images
364
Y.K. Park and J.K. Kim, “Fast adaptive smoothing based on LBP for robust face recognition”, IEE Electronics Letters, 2007, 43 (24):13501351. K. Mikolajczyk and C. Schmid, “Indexing based on scale invariant interest points”, in: 8th IEEE International Conference on Computer Vision, 2001, pp. 525-531. M. Heikkilä and M. Pietikäinen, “A texture-based method for modeling the background and detecting moving objects”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2006, 28(4):657-662. T. Ojala, M. Pietikäinen, and D. Hardwood, “A comparative study of texture measures with classification based on feature distribution”, Pattern Recognition, 1996, 29:51-59. T. Ojala, M. Pietikäinen, and T. Mäenpää, “Multiresolution grayscale and rotation invariant texture classification with Local Binary Patterns”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2002, 24(7):971- 987. M. Heikkilä, M. Pietikäinen, and C. Schmid, “Description of interest regions with local binary patterns”, Pattern Recognition, 2009,42(3):425-436. T. Mäenpää and M. Pietikäinen, “Multi-scale binary patterns for texture analysis”, Lecture Notes in Computer Science 2749, pp. 885892. W.Y. Liu, Z. Su, S. Li, et al, “A performance evaluation protocol for content-based image retrieval algorithms/systems”, In CD-ROM of IEEE CVPR Workshop on Empirical Evaluation Methods in Computer Vision, Hawaii, USA. L.Y. Ru, X. Peng, Z. Su, et al, “Feature performance evaluation in content-based image retrieval”, Journal of computer research and development, 2003, 40(11):1566-1560. (In Chinese)
