|
One of the main goals of texture analysis is to provide a robust mathematical description of the spatial behavior
of intensity values in any given neighborhood. These local distributions {called textures{ characterize object
surfaces and are used for pattern identication and recognition of images. However, some spatial patterns may
vary from quite simple stripes to randomness, where textures look like unstructured noise. Since textures can
exhibit a large number of properties such as surface materials and geometry of the lighting sources, many dierent
approaches have been proposed. A featured method is the modication of Wang's algorithm made by Ojala et
al, the so-called local binary patterns (LBP). The LBP algorithm uses a 3×3 square mask named "texture
spectrum" which represents a neighborhood around a central pixel. The values in the square mask are compared
with the central pixel and then multiplied by a weighting function according with their positions. This technique
has become popular due to its computational simplicity and more importantly for encoding a powerful signature
for describing textures. Specially, it has gained increased importance in image classication, where the success
not only depends on a robust classier but also relies in a good selection of the feature descriptors. However,
Ojala's algorithm presents some limitations such as noise sensitivity and lack of invariance to rotational changes.
This fact has fostered many extensions of the original LBP approach that in many cases are based on minor
changes in order to attain e.g. illumination and rotational invariance or improving the robustness to noise. In
this paper we present a detailed overview of the LBP algorithm and other recently modications. In addition, we
perform a texture classication study with seven algorithms in presence of rotational changes, noise degradation,
contrast information, and dierent sizes of LBP masks using the USC-SIPI database. The LBP histograms have
been evaluated using the Kullback-Leibler distance. This study will be a valuable insight for establishing a robust
and ecient texture descriptor to solve real world problems.
|
