Gender Classification on Real-Life Faces - Advanced Concepts for Intelligent Vision Systems - page 323

Table 1. Overview of recent studies on gender classification from face images

Data Set Approach

Study Performance

Data Real-Life Public Feature Classifier

2002 [5] 1,755 No Yes raw pixels SVM 96.62%

2002 [10] 3,500 Yes No haar-like features Adaboost 79.0%

2005 [6] 12,964 No Yes local-region matching SVM 94.2%

2006 [7] 5,326 No Yes fragment-based boosting 91.72%

filter banks

2007 [8] 2,409 No Yes pixel comparisons Adaboost 94.3%

2008 [9] 500 No Yes raw pixels SVM 86.54%

2009 [11] 10,100 Yes No haar-like features probabilistic 95.51%

boosting tree

our work 7,443 Yes Yes boosted LBP features SVM 94.44%

(including FERET and PIE), local region-based SVM achieved the performance
of 94.2%. Lapedriza et al. [7] compared facial features from internal zone (eyes,
nose, and mouth) and external zone (hair, chin, and ears). Their experiments on
the FRGC database show that the external face zone contributes useful infor-
mation for gender classification. Baluja and Rowley [8] introduced an ecient
gender recognition system by boosting pixel comparisons in face images. On the
FERET database, their approach matches SVM with 500 comparison operations
on 20

20 pixel images. Makinen and Raisamo [9] systematically evaluated dif-
ferent face alignment and gender recognition methods on the FERET database.

×

Fig. 1. Examples of real-life faces (from the LFW database). ( top 2 rows ) Female;
( bottom 2 rows )Male.