Transcript Document 7273133

Automatic Face Recognition
Using Color Based
Segmentation and Intelligent
Energy Detection
Michael Padilla and Zihong Fan
Group 16
EE368, Spring 2002-2003
Project Objective
Given a digital image of attractive and intelligent EE368
students and teaching staff, detect the presence of faces
in the image and output their location and (if poss.) gender.
Basic System Summary
• Initial Design




Reduced Eigenface-based coordinate system defining a “face space”, each possible face a point in space.
Using training images, find coordinates of faces/non-faces, and train a neural net classifier.
Abandoned due to problems with neural network: lack of transparency, poor generalization.
Replaced with our secondary design strategy:
• Final System
Input
Image
Color-space
Based
Segmentation
Morphological
Image
Processing
Matched
Filtering
Peak/Face
Detector
Face
Estimates
H vs. S vs. V (Face vs. Non-Face)
For faces, the Hue value is seen to typically occupy values in the range
H < 19
H > 240
We use this fact to remove some of the non-faces pixels in the image.
Y vs. Cr vs. Cb
In the same manner, we found empirically that for the YCbCr space
that the face pixels occupied the range
102 < Cb < 128
125 < Cr < 160
Any other pixels were assumed non-face and removed.
R vs. G vs. B
Finally, we found some useful trends in the RGB space as well. The
Following rules were used to further isolate face candidates:
0.836·G – 14 < B < 0.836·G + 44
0.89·G – 67 < B < 0.89·G + 42
Removal of Lower Region – Attempt to
Avoid Possible False Detections
Just as we used information regarding face color, orientation, and scale from
The training images, we also allowed ourselves to make the assumption that
Faces were unlikely to appear in the lower portion of the visual field: We
Removed that region to help reduce the possibility of false detections.
Morphological Processing
Step 1: Open Operation
After removing pixels based on color space considerations,
removed specs initially by use of the open operation with
a window of size 3x3.
Morphological Processing
Step 2: Small “Blob” Removal
Model the average size of head blobs in the training
reference image. Remove blobs below one standard deviation.
In addition, we:
• Convert to grayscale. In our case, no more color information to extract.
• Apply mean removal+histogram equalization -> flatten and bring out details.
Template Design
• Manually selected a number of quality faces: centered, straight, lighting, diverse.
• Measured face dimensions and used Matlab to uniformly scale and align them.
• Efforts resulted in 26 sample faces added to produce the final template.
Final Face
Template
Original Faces
Matched Filter Operation
Scale and Rotate
Pre-processing
Masked Input
Image
Apply
Matched Filter
While (remaining mask area to analyze) {
for s = 1:S { % scale
for r = 1:R { % rotation
for thrshld = Max:Min
template = temp(mother_temp, s, r);
peaks = conv(mask_image, template);
face = detector(peaks, thrshld);
if (face) {
adjust mask_image;
adjust remaining mask area;
}
}
}
}
}
Compare peaks
To Threshold,
T(n)
If
Peak > Tn,
Declare face
For each scale and rotation,
the threshold, T(n), decreases
When faces are detected, we remove the corresponding portion of the
masked input image to try to avoid multiple and false detections
Algorithm is sensitive to errors made in the pre-processing stage.
Face
Coordinates
Matched Filtering - Steps
Pre-processing
Masked Input
Image
Apply
Matched Filter
Compare peaks
To Threshold,
T(n)
If
Peak > Tn,
Declare face
Face
Coordinates
Face Detection Steps and
Progressive Masking
After detecting peaks at the output
of the matched filter, the following
steps are taken:
• Peaks within threshold range -> faces.
• Face pixels are convolved with oval
face mask of appropriate scale.
Removes neighborhood of detected
face pixel.
• After all processing, face pixels are
consolidated into blobs by dilation.
• Finally, centroids of blobs deemed to
be face centers.
General Results
(Example result for Training_7.jpg)
•
•
•
•
For training images, run time 80 – 110 sec.
Detection results range from 83% - 100%.
Main Strengths: Intuitive and (thus far) accurate.
Main Weaknesses: Sensitive to errors in pre-processing.
Conclusions
• In most cases, effective use of color space – face color
relationships and morphological processing allowed
effective pre-processing.
• For images trained on, able to detect faces with reasonable
accuracy and miss and false alarm rates.
• Adaptive adjustment of template scale, angle, and threshold
allowed most faces to be detected.
• Decision Feedback Masking reduced multiple and false
detection rate
If additional time, would have liked to:
• Pursue the Eigenimage approach further with MRC or SVM.
• Explore use of Wavelet spaces for face/gender detection.
References
• Bernd Girod, EE368 Class Lecture Notes, Spring 2002-2003
• R. Gonzalez and R. Woods, “Digital Image Processing – 2nd
Edition”, Prentice Hall, 2002
• C. Garcia et al., “Face Detection in Color Images Using
Wavelet Packet Analysis”.
• M. Elad et al., “Rejection Based Classifier for Face
Detection”, Pattern Recognition Letters, V.23, 2002.
Face Detection
Gender Recognition
1
5 (19)
2 (1)
2
3 (20)
5 (0)
3
10 (18)
5 (0)
4
10 (18)
5 (0)
5
15 (13)
5 (0)
6
12 (17)
5 (0)
7
13 (16)
5 (0)
8
14 (14)
1 (2)
9
1 (22)
2 (1)
10
3 (20)
2 (1)
11
16 (08)
5 (0)
12
5 (19)
5 (0)
13
5 (19)
5 (0)
14
5 (19)
5 (0)
15
5 (19)
5 (0)
16
2 (21)
5 (0)
17