Transcript LDP

LDP
Local Directional Pattern &
LDN
Local Directional Number Pattern
报告人:黄倩颖
内容
两种局部编码模式构造描述子
LDP Local Directional Pattern
LDN Local Directional Number Pattern
对Local Binary Pattern (LBP)的改良
Descriptor
geometric-feature-based
appearance-based
Part One
作者简介
文章结构
方法概述
讲解提纲
•
•
•
•
•
•
LBP方法回顾
LDP的创新点
LDP的鲁棒性
LDP的旋转不变性
实验
结论
作者简介
Local Directional Pattern (LDP) – A Robust
Image Descriptor for Object Recognition
Taskeed Jabid, Md. Hasanul Kabir, Oksam Chae
Department of Computer Engineering Kyung Hee University,
Republic of Korea
2010 Seventh IEEE International Conference on Advanced Video and
Signal Based Surveillance
Taskeed Jabid
Human Computer Interaction, Computer Vision,
Object Recognition
Local Directional Pattern (LDP) for face recognition
International Conference Consumer Electronics (ICCE), 2010
Cited by 44
文章结构
• Introduction
• LDP image descriptor
a.
b.
c.
d.
e.
Local Binary Pattern (LBP)
Local Directional Pattern (LDP)
Robustness of LDP
Rotation invariant LDP
LDP Descriptor
• Texture classification using LDP descriptor
• Face recognition using LDP descriptor
• Conclusions
Abstract
LDP( Local Directional Pattern) is
a local feature descriptor for describing local
image feature.
• Though LBP is robust to monotonic illumination
change but it is sensitive to non-monotonic
illumination variation and also shows poor
performance in the presence of random noise
• A LDP feature is obtained by computing the edge
response values in all eight directions at each pixel
position and generating a code from the relative
strength magnitude. Each bit of code sequence is
determined by considering a local neighborhood
hence becomes robust in noisy situation.
Part One
作者简介
文章结构
方法概述
讲解提纲
•
•
•
•
•
•
LBP方法回顾
LDP的创新点
LDP的鲁棒性
LDP的旋转不变性
实验
结论
讲解提纲
•
•
•
•
•
•
LBP方法回顾
LDP的创新点
LDP的鲁棒性
LDP的旋转不变性
实验
结论
Local Binary Pattern (LBP)
Original LBP
26 < 50
0
1
85
32
26
53
50
10
1
60
38
45
1
Threshold 50
0
0
0
0
(0 0 1 1 1 0 0 0)2 = 56
0
Local Directional Pattern (LDP)
Kirsch masks
NorthWest
West
SouthWest
North
M 3 M2 M1
5
5
-3
5
5
5
-3
5
5
5
0
-3
-3
0
-3
-3
0
5
-3
-3
-3
-3
-3
-3
-3
-3
-3
M2
M1
M4
5
-3
-3
M3
5
0
-3
M4
5
-3
-3
M5
M0
399
M6
M7
M0
85
-3 32
-3 26
5
53
-3 50
0 10
5
East
60
-3 38
-3 45
5
M 5 M6 M7
-3
-3
-3
-3
-3
-3
-3
-3
-3
5
0
-3
-3
0
-3
-3
0
5
5
5
-3
5
5
5
-3
5
5
South
NorthEast
SouthEast
Computing…
85
32
26
313
97
503
Kirsch masks
53
50
19
10
537
60
38
45
161
0
LDP Binary Code =
00010011
1
LDP Decimal Code=
19
0
0
1
1
0
0
399
97
161
LDPk
k=3
Robustness of LDP
noise & non-monotonic illumination changes
85
81
-4
32
29
-3
26
32
-6
85
32
26
53
50
10
53
50
10
38
58
15
-15
+8
+5
60
38
45
60
38
45
65
43
47
+5
+5
+2
LBP = 00111000
LDP = 00010011
LBP = 00101000
LDP = 00010011
Rotation invariant LDP
85
32
26
53
50
10
60
38
45
26
10
45
32
50
38
85
53
60
313
97
537
503
399
161
97
161
503
393
161
97
313
97
537
161
0
0
1
1
1
0
0
0
1
1
0
0
0
0
1
Rotation Invariant LDP Code = 0 0 1 1 0 0 0 1
0
LDP Descriptor
Accumulating the occurrence of LDP feature
Experiments
Texture Classification using LDP histogram
Primary pictures
from Brodatz
texture album:
(a) Bark,
(b) Brick,
(c) Bubbles,
(d) Grass,
(e) Leather,
(f) Pigskin,
(g) Raffia,
(h) Sand,
(i) Straw,
(j) Water,
(k) Weave,
(l) Wood and
(m) Wool
Experiments
Texture Classification using LDP histogram
Experiments
Extracted rotation invariant LDP features of each
pixel of the image then combined to generate
rotation invariant image descriptor using LDP
histogram following equation.
Experiment Results
The accuracy of the method
Results
Face recognition using LDP descriptor
Database FERET
(a) fa set, used as a gallery set, contains frontal images of 1,196 people.
(b) fb set (1,195 images) with an alternative facial expression than in
the fa photograph.
(c) fc set (194 images) taken under different lighting conditions.
(d) dup I set (722 images) taken later in time.
(e) dup II set (234 images) subset of the dup I set containing images
that were taken at least a year after the corresponding gallery image.
Face recognition using LDP descriptor
Classification using LDP histogram
Template matching
Experiment Results
Part Two
作者简介
文章结构
方法概述
讲解提纲
•
•
•
•
•
LBP LDP缺点
LDN 三个关键点
人脸描述
实验
结论及未来工作
作者简介
Local Directional Number Pattern for Face
Analysis: Face and Expression Recognition
Adin Ramirez Rivera,Student Member, IEEE,
Jorge Rojas Castillo,Student Member, IEEE,
and Oksam Chae,Member, IEEE
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 22, NO. 5, MAY 2013
Cited by 2 | Year 2012 |
Adin Ramirez Rivera
Image Processing, Computer Vision
Content-Aware Dark Image Enhancement through Channel Division
IEEE Transactions on Image Processing 21 (9), 3967-3980
Cited by 9 | Year 2012
文章结构
• Introduction
• Local Directional Number Pattern
• Difference With Previous Work
• Coding Scheme
• Compass Masks
• Face description
• Face recognition
• Conclusions
Abstract
A novel local feature descriptor
LDN encodes the directional
information of the face’s textures in
a compact way, producing a more
discriminative code than current
methods
Part Two
作者简介
文章结构
方法概述
讲解提纲
•
•
•
•
•
LBP LDP缺点
LDN 三个关键点
人脸描述
实验
结论及未来工作
讲解提纲
• LBP LDP缺点
• LDN 三个关键点
• 人脸描述
• 实验
• 结论及未来工作
LBP
The method discards most of the
information in the neighborhood.
It limits the accuracy of the method
It makes the method very sensitive to noise
Moreover, these drawbacks are more evident for
bigger neighborhoods
Directional (LDiP) & Derivative (LDeP)
Miss some directional information
(the responses’ sign) by treating all
directions equally
Sensitive to illumination changes and noise, as the
bits in the code will flip and the code will represent
a totally different characteristic
Key points of LDN
Direction
number
Sign
information
LDNLBP
6-bit
gradient
information
Key points of LDN
LDN
Sign
Direction
number information
6-bit
gradient
information
Coding Scheme
Direction
number
+ Sign -
information
Coding Scheme
Compass Masks
Two kinds of masks
𝐿𝐷𝑁
𝐾
Kirsch masks
𝐿𝐷𝑁𝜎𝐺 derivative-Gaussian mask
Compass Masks
Kirsch masks
NorthWest
West
SouthWest
North
M 3 M2 M1
5
5
-3
5
5
5
-3
5
5
5
0
-3
-3
0
-3
-3
0
5
-3
-3
-3
-3
-3
-3
-3
-3
-3
M2
M1
M4
5
-3
-3
M3
5
0
-3
M4
5
-3
-3
M5
M0
M6
M7
M0
-3
-3
5
-3
0
5
-3
-3
5
M 5 M6 M7
-3
-3
-3
-3
-3
-3
-3
-3
-3
5
0
-3
-3
0
-3
-3
0
5
5
5
-3
5
5
5
-3
5
5
South
NorthEast
East
SouthEast
Compass Masks
derivative-Gaussian mask
• Compute code in gradient space
• Therefore, use Gaussian smoothing to stabilize the
code in presence of noise
Generate a compass mask,{M0σ,...,M7σ}, by rotating Mσ,
45°apart, in eight different directions
Compass Masks
derivative-Gaussian mask
Face Descriptor
Histogram
LH & MLH
Face Descriptor
Two kinds of descriptor
Code in LH
Code in MLH must be
Face Recognition
Chi-Square dissimilarity measure
Face recognition using LDP descriptor
Database FERET
(a) fa set, used as a gallery set, contains frontal images of 1,196 people.
(b) fb set (1,195 images) with an alternative facial expression than in
the fa photograph.
(c) fc set (194 images) taken under different lighting conditions.
(d) dup I set (722 images) taken later in time.
(e) dup II set (234 images) subset of the dup I set containing images
that were taken at least a year after the corresponding gallery image.
Experiment Results
Face recognition accuracy
small neighborhoods (3×3, 5×5, 7×7)
medium neighborhoods (5×5, 7×7, 9×9)
large neighborhoods (7×7, 9×9, 11×11)
Experiment Results
Noise Evaluation
With white Gaussian noise
Conclusion
• Combination of different sizes (small, medium
and large) gives better recognition rates for
certain conditions.
• Evaluated LDN under expression, time lapse
and illumination variations, and found that it is
reliable and robust throughout all these
conditions.
总结及未来工作
• 如何选择一个描述子
•
•
•
•
长度
描述精度
抗噪能力
计算强度
• 如何设计一个描述子
• 舍弃冗余的信息
• 整合多种信息来源
• 信息压缩