Implementing the local binary patterns with SIMD instructions of CPU

Abstract

Usage of statistical classifiers, namely AdaBoost and its modifications, is very common in object detection and pattern recognition. Performance of such classifiers strongly depends on low level features they use. This paper presents an experimental implementation of the Local Binary Patterns (LBP) that uses SIMD instructions for acceleration. The experiments shows that the proposed implementation is about six times faster than the plain C implementation (i.e. with no special optimizations) and superior to optimized implementations of features with similar descriptive power.