Information Technology Reference
In-Depth Information
13.5 Experimental Study
A set of experiments were conducted in order to study the impact of the selected
moment subsets on the overall recognition performance in several pattern recognition
problems. For this purpose, appropriate software was developed in the MATLAB
2012b environment, while all experiments were executed in an Intel i5 3.3GHz
PC with 8GB RAM. Moreover, four well known benchmark datasets were used
to evaluate the initial assertion of the moments' selection significance, towards the
improvement of the recognition accuracy.
13.5.1 Benchmark Datasets
Four benchmark datasets are used in order to investigate the selection performance
of the Relief and Genetic Algorithms in selecting moment subsets of various sizes.
The considered datasets are the Yale face recognition dataset [
1
], a subset [
22
]of
the Terravic thermal infrared face recognition [
18
], the JAFFE [
17
] and RADBOUD
[
16
] facial expressions datasets. Three sample images (different classes) from each
dataset are depicted in Fig.
13.9
, while the properties of each dataset are summarized
in Table
13.1
.
13.5.2 Datasets Pre-processing
It is worth noting that before computing the moment features, the images need to
be pre-processed in order to remove irrelevant image information (background, hair,
ears, etc.) and to isolate the image's part, which includes the main face information.
For this purpose, the Viola-Jones face detector [
29
] is applied being followed by
an ellipse masking [
11
], for the case of Yale, JAFFE and Radboud datasets. The
aforementioned face detector fails to detect faces in thermal infrared images and
Table 13.1
Benchmark datasets properties
Dataset
Type
Number of classes Samples/class
Total samples
Ya l e
Face recognition
15
11
165
Terravic
Thermal infrared face
recognition
10
70
700
JAFFE
Facial expression
recognition
7
30, 29, 32, 31, 30, 31, 30 213
Radboud -
8
67
536
Search WWH ::
Custom Search