Selection of Visual Descriptors for the Purpose of Multi-camera Object Re-identification - Feature Selection for Data and Pattern Recognition - page 268

(a)

(b) (c) (d) (e) (f)

° °

vertical horizontal diagonal 45 diagonal 135 none

Fig. 12.8 Edge Histogram Descriptor: a Consecutive stages of the image segmentation; b - f direc-
tions for each bin in the edge histogram

The acquired maximum value is additionally thresholded to remove weak direc-
tionalities which could be caused by image noise:
max m [ dir ] >

T edge (12.11)

The applied threshold ( T edge ) is set to 10 during experiments. If this condition is
fulfilled, the appropriate histogram bin value, related to the specified directionality, is
incremented. In this way, 5 bin histograms for each of the 16 image regions are built.
Hence, the feature vector acquired utilizing this method consists of 80 elements.

12.4.7 Local Binary Pattern

Local Binary Pattern (LBP) is a texture related parametrization method which
describes neighbourhood of each analysed image point [ 11 , 24 ]. In most cases, a
direct 8-point neighbourhood of an analysed pixel is used. However, it is possible to
define it in other words. LBP extraction procedure can be divided into two stages.
First, each pixel neighbourhood is analysed to determine the related code word. This
is done by comparing neighbouring pixel values against the analysed pixel. For the
pixel positions where the analysed point is greater than the related neighbour, the
value of 1 is assigned, 0 otherwise. In this way, binary code words are formed for each
of the image points. Afterwards, they are used to build a histogram. The description
obtained with this approach is rotation-dependent. Therefore, before feature extrac-
tion occurs, object images are preprocessed to set their proper orientation.
The LBP histogram is calculated and normalized for each of the colour channels
independently. After quantizing each of them to 64 bins, a feature vector consisting
of 192 elements is built ( LBPHist descriptor).