Gender Classification on Real-Life Faces - Advanced Concepts for Intelligent Vision Systems - page 326

Fig. 4. Each face image is divided into sub-regions from which LBP histograms are
extracted and concatenated into a single, spatially enhanced feature histogram

In [15], Adaboost was used to select the discriminative sub-regions (in terms
of LBP histogram) from a large pool generated by shifting and scaling a sub-
window over face images. In contrast, here we look at regional LBP histograms
at the bin level, to identify the discriminative LBPH bins.

3 Experiments

All experimental results were obtained using the 5-fold cross-validation. We parti-
tioned the data set into five subsets of similar size, with a similar balance between
the two classes. The images of a particular subject appear only in one subset. In
each trial, one subset was used for testing, while the remaining four subsets were
used for training. The recognition results were averaged over the 5 trials.
In our experiments, we used the SVM implementation in the library SPIDER 1 .
The Radius Basis Function (RBF) kernel was utilized, and the parameters were
tuned to obtain the best performance. Meanwhile, each dimension of the feature
vector was scaled to be between -1 and 1. As a baseline to compare against, we
also applied SVM with raw image pixels, which delivers the best performance
on face images acquired in controlled environments [5]. For computational sim-
plicity, face images of 127 46 pixels, thus

each image represented by a vector of 2,944 dimensions. We summarize the re-
sults of SVM with raw pixels and standard LBP features in Table 2. As can
be observed, LBP features produce better performance than raw image pixels.
Regarding support vectors, with raw pixels, the learned SVMs utilized 51-53%
of the total number of training samples (in each trial of cross-validation, the
number varies slightly), while SVMs with LBP features employ 58-61%.
For boosting learning, to generate a large LBP feature pool, we can generate
many more sub-regions by shifting and scaling a sub-window over face images.
In this study, we fixed the size of sub-window as 18

91 pixels were down-scaled to 64

× ×

15 pixels, and shifted the
sub-window with the shifting step of 4 pixels vertically and 3 pixels horizontally.
In total 700 sub-regions were obtained. By applying 59-label LBP (8 , 2 ,u 2) to
each sub-regions, a histogram of 413,000 (700

×

59) bins was extracted from each
face image. We adopted Adaboost to learn discriminative LBPH bins and boost

×

1 http://www.kyb.tuebingen.mpg.de/bs/people/spider/index.html