Transcript Statistical Evaluation of Detection Results - WWW2

逄霖生
中國文化大學 電機工程學系
Outline

Introduction
Statistical Detection Models








Acquisition of Human Face Images
Skin Detection
Ocular Region Detection
Experimental Results
ROC Evaluation
Conclusions
Future Works
Introduction
 To Detect human face by using statistical methods
 The given image is treated as an random variable.
 The colors and other features of data is treated as the
outcomes of the given random variable.
 The prior and posterior information can be used to
handle statistical data.
 The uncertainty of information reveals the variations of
data.
 The ROC curve statistically evaluates the detection
results.
Statistical Detection Models
• Bayes’ Filter for Skin Detection
• Entropy Model for Eye Detection
• ROC (Receiver Operating Characteristic) curve
for statistical evaluation of skin detection results
Statistical Detection Models
• Bayes Rule
p( x  y ) p ( y | x) p ( x)
P( x | y ) 

p( y )
p( y )
• Entropy
N
   p( I i )(log 2 p( I i ))
i 1
• ROC curve (Receiver Operating Characteristic curve)
• Statistical Evaluation of Detection Results
Flow Chart
Raw Image Data
Skin Detection
(Bayes filter)
Eye & Eyebrow
Detection
(Entropy analysis)
Performance
Evaluation
(ROC curve)
Color
Conversion
Mouth Detection
(Color ratio analysis)
Face Detection
Color Space Conversion
1. RGB
Primary colors (tri-stimulus values of
colors)
2. YCbCr
Luminance & Chrominance
 Y   16   65.481 128.553 24.966   R 
C   128   37.797  74.203 112.000  G 
 b   
 
Cr  128  112.000  93.786  18.214  B 
s ＝ T(r)
where “s” is an output image, “r” is an input
image
3. Gray Level
Image Acquisition
[Left] Original Image, [Right] Pre-selected Skin Area
Note that the eye & eye brow, mouth are not part of skin
Skin Detection
• Applying a Bayes Filter to an image
p( x  y ) p ( y | x) p ( x)
P( x | y ) 

p( y )
p( y )
where
p(x) and p(y) are pdfs of random variables x and y,
p(x|y) is the posterior probability
p(y|x) is the prior probability.
Skin Detection
p(x|y)
0.4
Cb-Skin
Cr-Skin
Cb-NSkin
Cr-NSkin
0.35
Prob. (normalized)
0.3
0.25
0.2
P( x | y ) 
0.15
p ( y | x) p( x)
p( y )
0.1
0.05
0
0
50
100
150
Color (x=[0..255])
200
250
Skin Detection
 By using a Bayes filter and a thresholding method, the
skin detection result of an image is shown as follow:
Skin Detection
Morphology
Entropy
• Entropy(熵)
N
   p( I i )(log 2 p( I i ))
i 1
p(Ii) is the probability for the outcome Ii
 Measure the degrees of uncertainty for different
outcomes from a given random event
Ocular Region Detection
M
 r   p( I ir ) log 2 ( p( I ir )) , r  1.2...N .
i 1
N
 c   p( I ij ) log 2 ( p( I cj )) , c  1.2...M.
r
j 1

c
Lip/Mouth Detection
1,
M1  
0,
if t2  G
otherwise
R
 t1
.

1, if t 4  B R  t3
M2  
.

0, otherwise
Experimental Result
ROC Curve
• ROC Curve (Receiver Operating Characteristic curve)
• ROC analysis provides tools to select possibly optimal
models and to discard suboptimal ones independently
from the cost context or the class distribution.
• ROC analysis is related in a direct and natural way to
cost/benefit analysis of diagnostic decision or quality
making.
• It is widely used in binary discrimination evaluation.
Evaluation of Skin Detection
TPP
True Positive Possibility =sensitivity
FNP
False Negative Possibility
FPP
False Positive Possibility =1-specificity
TNP
True Negative Possibility
p(x|y)
0.4
0.35
0.3
1
TPP(Cr)
FPP(Cr)
TPP(Cb)
FPP(Cb)
0.9
0.8
0.7
TPP & FPP
0.25
0.2
0.15
0.1
0.6
0.5
0.4
0.3
0.2
0.05
0.1
0
0
50
100
150
Color (x=[0..255])
200
250
0
0
50
100
ROC
Cr
Cb
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
150
Color (x=[0..255])
1
TPP(sensitivity)
Prob. (normalized)
TPP(sensitivity) & FPP(1-specificity)
Cb-Skin
Cr-Skin
Cb-NSkin
Cr-NSkin
0.6
FPP(1-specificity)
0.7
0.8
0.9
1
200
250
ROC
Curve
Non-Skin indeterminate
Area
Area
Skin
Area
Cr
FPP(cr)<0.96
0.96≦FPP(cr)
TPP(cr)>0.9
Cb
FPP(cb)>0.02
0.02≦FPP(cb)≦0.08
TPP(cb)<0.025
Conclusions
 Statistical methods are able to classify and detect
human characteristics.
 Using the prior information can help us to recognize
the posterior situation.
 The uncertainty of analyzed data gives the location of
the area of eye.
 ROC curve can determine the content of experimental
results.
Future Works
 Adapted with Environmental Variations
 Hardware Acceleration