1. Face Detection EE368 Final Project Group 14 Ping Hsin Lee
Vivek Srinivasan
Arvind Sundararajan

2. Overview Introduction Methods used to detect faces
Color segmentation
Morphological Processing
Template Matching And Clustering
Results
Techniques considered but not used

3. Color Segmentation Use color information in the YCbCr domain
YCbCr Color space effectively decorrelates the intensity and color information
Each channel information is represented in discrete levels.

4. MAP Rule Implement MAP decoder to determine skin from non-skin pixels
D(I(x,y)) = 1 if P(I(x,y) | S)P(S) > T* P(I(x,y) | NS)P(NS)
= 0 other wise
Minimizes misclassification error

5. Result of Color Segmentation

6. Morphological Processing
Reject blobs of small sizes, perform closing, remove holes

7. Non-face Object Removal
Use information about shape and location of objects in conjunction to reject non-face objects while minimizing rejection of faces
Objects characterized by max/min as a measure of length. (Independent of size, translation, and rotation of objects)
Example of non-face object removed by CCA

8. Non-face Object Removal
Before and after rejection

9. Template Matching Performed in the luminance domain using the FFT
First attempt: use the average of all face regions
Features did not seem to align properly, hence this template was rejected
Rejected Template

10. Final Templates Used
Resample each face region to the same size before averaging.
Include mirror images of each face region to produce a symmetric template (a).
In addition, a non-symmetric partial template (b) is used to capture information about smaller and partially obscured faces in the image
One template tests for symmetry, while the other tests for non-uniform illumination, and captures smaller faces as well.

11. Clustering of Correlation Peaks
The autocorrelation results for each template were first thresholded and then combined.
Used heuristic techniques based on shape of the skin regions to group peaks.
Any 2 peaks meeting a maximum distance criterion and connected by a line passing through only skin regions were grouped together as a face.

12. Results of Grouping Correlation Peaks
Before and after peak grouping

13. Results Applied to the Original Image
Image corresponding to the grouped peaks

14. Final Results
Image 1 2 3 4 5 6 7
Score 20 22 24 21
Total 25

15. Techniques Considered but not Used
Fisher’s linear discriminant (FLD)
Poor performance in rejection of false positives because detected non-face and face regions are not linearly separable
Eigenfaces
Produced results similar to template matching but at an increased computational cost

16. Techniques Considered but not Used
Support Vector Machines (SVM)
Generated 470 face regions and 500 non-face regions each of size 49x55 pixels as training database
Employed a Gaussian radial basis function (RBF) as kernel

17. Samples of database images
Faces
Non-faces

18. Results of SVM
Produced decision regions that are too tightly bound to the training face samples and were not able to classify the faces in the other training pictures
Including the SVM in the program would only slow down our runtime and would not produce noticeable improvements

19. Conclusion
Color segmentation in the YCbCr domain and morphological processing produced good estimates of face regions
Implemented multi-resolution template matching and peak clustering to further distinguish different face regions from each other and from non-face regions
Could have done more to reject false positives (MRC/neural networks to reject hand regions)