Avoiding the Curse of Dimensionality in Local Binary Patterns

Local Binary Patterns is a popular grayscale texture operator used in computer vision for classifying textures. The output of the operator is a bit string of a defined length, usually 8, 16 or 24 bits, describing local texture features. We focus on the problem of succinctly representing the patterns using alternative means and compressing them to reduce the number of dimensions. These reductions lead to simpler connections of Local Binary Patterns with machine learning algorithms such as neural networks or support vector machines, improve computation speed and simplify information retrieval from images. We study the distribution of Local Binary Patterns in 100000 natural images and show the advantages of our reduction technique by comparing it to existing algorithms developed by Ojala et al. We have also confirmed Ojala's findings about the uniform LBP proportions.