A simple method for improving local binary patterns by considering non-uniform patterns

The basic idea behind LBP is that an image is composed of micropatterns. A histogram of these micropatterns contains information about the local features in an image. These micropatterns can be divided into two types: uniform and non-uniform. In standard applications using LBP, only the uniform patterns are used. The non-uniform patterns are considered in only a single bin of the histogram that is used to extract features in the classification stage. Non-uniform patterns have undesirable characteristics: they are of a high dimension, partially correlated, and introduce unwanted noise. To offset these disadvantages, we explore using random subspace, well-known to work well with noise and correlated features, to train features based also on non-uniform patterns. We find that a stand-alone support vector machine performs best with the uniform patterns and random subspace with histograms of 50 bins performs best with the non-uniform patterns. Superior results are obtained when the two are combined. Based on extensive experiments conducted in several domains using several benchmark databases, it is our conclusion that non-uniform patterns improve classifier performance.

► To use also non-uniform patterns when the LBP texture descriptor is used for describing a given image. ► The non-uniform patterns have undesirable characteristics: high dimension; correlation; noise. ► To offset the disadvantages of non-uniform patterns we use the random subspace. ► To build an ensemble considering both uniform and non-uniform patterns.