The Effect of Distance Function for NN Classifier with Local Binary Pattern Descriptors

The aim of pattern classification is to put similar patterns into the same cluster. Hence most classifiers employ distance functions or dissimilarity function to measure the dissimilarity of patterns. Some features are obtained by measuring the value extracted from the physical features of patterns directly. On the other hand, some features are obtained by counting the number of pattern property observed. Different distance functions will affect the performance of a classifier while different type of features are adopted. Recently, Local Binary Patterns (LBP) descriptor and its variants are the most popular methods for extracting features of detail textures of image patterns. Each LBP feature represents the observed number of a specific code. The Nearest Neighbor (NN) classifier is a simplest and efficient classifier to classify histograms of such pattern features. Here the effects of different distance functions for the NN classifier based on LBP features are studied. The UIUC texture database is chosen for observation. The experimental results reveal that the NN classifier using Hellinger distance function as the dissimilarity function to classify histograms of FbLBP features provides the best performance of classification with 92.4% and 94.1% accuracy under different radii.