EE368 Digital Image Processing Face Detection Project

							EE368 Digital Image Processing
   Face Detection Project


             By
      Gaurav Srivastava
       Siddharth Joshi
    Problem Definition


   To detect faces in a class group
    photograph.

   To differentiate female faces.
Challenges
   Varying lighting conditions.
   Various objects with pseudo-skin color.
   Occluded faces.
   Different scale size of faces.
   Faces in non-frontal position.
          Approach

Input
                                Morphological
Image       Skin Color                                        Eigenspace
                                  Operations
           Segmentation      (Hole Filling, Erosion)           Projection




           Detecting                                   Density Estimation
                               Deciding
          Male/Female                                         And
                             Face/Non-face
             Faces                                      Peak Detection
 Output
 Image


                    Block Diagram of Implementation
Skin Color Segmentation
   YCbCr Space
   Better Skin Color
    localization than
    HSV space.
   Invariant under
    various lighting
    conditions.
Result of Skin Color Segmentation
Morphological Operations
   Hole Filling.
   1st Level Erosion, Diamond structuring
    element.
   2nd & 3rd Level Column Erosion.
   Selection of blocks, by size criterion.
Binary Image After Hole Filling
Different Levels of Erosion
Eigenspace Decomposition
   Training set of 53
    facial images for KL
    Transform.
   First 20 eigenvectors
    used as Principal
    Components.
Gaussian F-space Density Estimation
   Estimation of the likelihood function for the
    image data – i.e. P(x|).

        P( x | )  PF x | PF x | 
        ˆ                     ˆ
   ˆ
    P( x | ) can be used to compute a local
    measure of the target saliency.
         (i, j ) ML  arg max{S (i, j; )}
                      
       S (i, j; )  P( x | )
Detected Face   Probability Density
RMS Detection Criterion
   Difference in
    reconstruction errors
    for Face/Non-face
    using eigenspace
    projections.
Gender Determination
   Projection calculations using multiple
    faces of a female.
   Calculation of RMSE of projections of a
    facial candidate with stored projections.
                  k
       MSE x   ( y x  yik ) ( y x  yik )
                               T

                 i 1

       MSE~  min( MSE x )  ~  Fi
          x                  x
Original Image
Detected Faces: Male/Female
Conclusions
   Combination of deterministic algorithms
    like PCA, F-space density estimation
    and heuristics.
   Difficult to generalize the algorithm.
   Algorithm performs well on most frontal
    faces.
   Difficulty in detecting occluded faces.
     Face Detection   Gender Recognition
1        2 (19)             2 (1)
2        1 (20)             3 (0)
3        3 (18)             3 (0)
4        3 (18)             3 (0)
5        8 (13)             3 (0)
6        5 (17)             3 (0)
7        6 (16)             3 (0)
8        7 (14)             1 (2)
9
10
11
12
13
14
15
16
17