Information Technology Reference
In-Depth Information
Table 2. Experimental results of gender classification
Approach
Recognition Rates (%)
Feature
Dim. Classifier Female Male
Overall
raw pixels
2,944
SVM
86.89 94.13 91.27 ± 1.67
standard LBP 2,478
SVM
89.78 95.73 93.38
1.50
±
boosted LBP
500 Adaboost 91.13 94.82 93.36 ± 1.49
boosted LBP
500
SVM
91.91 96.09 94.44
1.19
±
a strong classifier. We plot in Fig. 5 the average accuracy of Adaboost as a
function of the number of features selected. With the 500 selected LBPH bins,
Adaboost achieves recognition rate of 93.36%, which is comparable to that of
SVM using the standard LBP (2,478 bins). However, Adaboost is much more
computationally ecient than SVM, requiring much less features.
We plot in the left side of Fig. 6 the top 20 sub-regions that contain most
LBPH bins selected. The right side of Fig. 6 further shows the spatial distribution
of the selected 500 LBPH bins in the 5-fold cross-validation experiments, where
each small patch represents the corresponding sub-region, and the grayscale in-
tensity of each patch is proportional to the number of bins selected from that
sub-region. It is observed that the discriminative LBPH bins are mainly dis-
tributed in the regions around/above eyes. Although faces are (on average) sym-
metric, the selected features are not symmetric, because of the pose/illumination
variations in the dataset. Regarding the distribution of selected features among
59 bins, we plot in Fig. 7 the distribution of the 500 features selected. We can
observe that selected bins distribute in all 59 bins, but some bins do have more
contributions (e.g., bin 2, 12, 27, and 34).
We further adopted SVM with the selected LBPH bins for gender classifica-
tion, which achieves the best performance of 94.44%. Moreover, the numbers of
support vectors were 32-35% of the total number of training samples, which are
0.9
0.85
0.8
0.75
0.7
0.65
0
50
100
150
200
250
300
350
400
450
500
Number of Features
Fig. 5. Average recognition rate of Adaboost, as a function of the number of feature
used
Search WWH ::




Custom Search