next up previous contents
Next: Gradated Histogram Fitting Up: Shading and Lighting Normalization Previous: Windowing Histogram Analysis

Beards and Hair

Phillips and Vardi's technique [33] produced particularly poor results when the individual had a beard or other facial hair. The algorithm brightens the dark hair regions unnaturally. Consequently, the regions of the face that were likely to be obstructed with hair were omitted from the histogram generation process (both in the mean face and in new faces). Thus the inverse histogram mapping does not over-brighten the face to account for these extreme regions. The generation of Hl(i),Hr(i), $f_{H_l \rightarrow
G}$ and $f_{H_r \rightarrow G}$ was limited vertically to cover only the eyebrows down to the nose as shown in Figure [*]. However, we still apply the mappings in $f_{H_l \rightarrow
G}$ and $f_{H_r \rightarrow G}$ to the whole face as in Figure [*]. Thus, the illumination is corrected for quite nicely in the case of bearded individuals and other unusual cases.

Figure 4.19: Limiting histogram generation to avoid hair and beards.
\epsfig{file=norm/figs/,height=8cm, angle=-90} \end{figure}

Tony Jebara