当前位置:文档之家› 新的人脸识别特征集_A New Feature Set for Face Detection

新的人脸识别特征集_A New Feature Set for Face Detection


Type 1&2:


The red area ratio is only required to be [1/2 1) times the whole area The relative position of red area is not required to be in the middle of whole area. The relative position of red area is not required to be in the middle of white area. The width and height are not required to be the same, only even lengths are required.

TII type 2 (TII2) consists a right triangle with the right angle at the up-right corner. Let the width of the original image be W, TII2(1:W+1, -1) = 0, TII2(1:W+1, 0) = 0, (1, 1) TII2(0, y) = 0. TII2(x-1, y-2) TII2(x, y) = I (x, y) TII2(x-1, y-1) + TII2(x, y-1) I (x, y) + TII2(x-1, y-1) W TII2(x, y-1) - TII2(x-1, y-2).
Type 1
Type 2
Type 3
Type 4
MIR Lab, ISA/NTHU
Type 5
michael@.tw
Type 6
14
3.2.1 Extension of Rectangle Features (Cont.)

Extensions of each type:
0 0 0 0 0 0 0 0 0 0 0
0
0
0
0
0
0
0
0
0
0
0
0 0 0
0
0 0 0 0 0 0
W
MIR Lab, ISA/NTHU michael@.tw
TII1(x+1, y-1)
9
3.1.2 Triangle Integral Images (Cont.)
MIR Lab, ISA/NTHU michael@.tw 4
2. Related Work

2001 – Viola and Jones introduce:
Rectangle
integral image Training algorithm based on AdaBoost A cascaded structure
g. RII(x2+w2-1, y2-1) a. RII(x1-1, y1-1) h. RII(x2-1, y2-1) (x1, y1) (x2, y2) h2 f. TII1(x2+w2, y2-1) b. RII(x1+w1-1, y1-1) h1 w2 w1 d. RII(x1+w1-1, y1+h1-1)
MIR Lab, ISA/NTHU
michael@.tw
12
3.2 Features (Cont.)


A feature extracts a value by subtracting the sum of pixels of the light area from the sum of pixels of the dark area in an image. A sum of a rectangle area can be computed by four points in a RII:
2X

Three rectangle feature:
1
=
2
MIR Lab, ISA/NTHU
–2X
1
–2X
2
michael@.tw
16
3.2.2 Type 7: Triangle Feature Type A

This research introduces two new types of triangle features. Up-left right triangle feature (type 7):
a. RII(x-1, y-1) (x, y) height d. RII(x+width-1, y+height-1) b. RII(x+width-1, y-1)
c. RII(x-1, y+height-1)
width
The sum of pixels in gray area = d - b - c + a .
Image Resize image Contrast stretching
Transform to integral images
Face Detection
Faces Non-faces
MIR Lab, ISA/NTHU
michael@.tw
3
Outline
1. Introduction 2. Related Work 3. Integral Image and Features 4. Learning Strong Classifiers 5. Cascade of Classifiers 6. Experimental Results 7. Conclusions and Future Work
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
MIR Lab, ISA/NTHU
michael@.tw
10
3.2 Features

We can observe that a human face has some useful characteristics, for example:
I ( x' , y ' )
RII(x, y)
(x, y)

The RII at location (x, y) contains the sum of the pixels above and to the left of (x, y). We also calculate the square sum of the image calculating variance for contrast correction (see 6.1).
MIR Lab, ISA/NTHU
michael@.tw
7
3.1.1 Rectangle Integral Image

The definition of Rectangle Integral Image (RII):
RII( x, y)
x ' x , y ' y
MIR Lab, ISA/NTHU michael@.tw 2
1. Introduction

Face detection is an important component of a video information retrieval system. This research focuses on the upright-frontal face detection problem. System flow chart:
michael@.tw 13
MIR Lab, ISA/NTHU
3.2.1 Extension of Rectangle Features

We extend some feature types to be more generalized. Totally six types of rectangle features are u3, 4&5:


Type 6:

MIR Lab, ISA/NTHU
3.2.1 Extension of Rectangle Features (Cont.)

We calculate rectangle features by subtracting areas of two (type 1~5) or three (type 6) rectangles. Two rectangle feature: = –
michael@.tw 8
MIR Lab, ISA/NTHU
3.1.2 Triangle Integral Images

This research introduces two new types of triangle integral images (TIIs). TII type 1 (TII1) consists a right triangle with the right angle at the up-left corner. TII1(x+1, y-2) Let the width of the original image be W, TII1(1:W+1, -1) = 0, (1, 1) TII1(1:W+1, 0) = 0, TII1(W+1, y) = 0. TII1(x, y-1) TII1(x, y) = I (x, y) + TII1(x, y-1) I (x, y) + TII1(x+1, y-1) - TII1(x+1, y-2).
The
area of eyes is usually darker than the area of the bridge of the nose. The area across eyes is usually darker than the area of cheeks. (proposed by Viola)
相关主题