Evaluation of LBP variants using several metrics and kNN classifiers

Abstract

In this paper, we demonstrate that the Adaptive Local Binary Pattern with oriented Standard deviation (ALBPS) method outperforms the original local binary pattern (LBP) as well as some of its most recent variants: Adaptive Local Binary Pattern (ALBP), Complete Local Binary Pattern (CLBP) and Local Binary Pattern Variance (LBPV). All the descriptors have been tested using two different dataset, KTH-TIPS 2a, a challenging multiclass dataset for material recognition and a binary sperm dataset for vitality classification. Three variants of the non parametric method of nearest neighbours combined with four metric distances have been used in the retrieval step in order to draw a more decisive conclusion. Best results were achieved when describing the images with ALBPS in both datasets. In regard to the KTH-TIPS 2a, the best performance is obtained using the weighted kNN with a 61.47% of hit rate using ALBPS and Chi Square distance, outperforming the ALBP in 1,07% and the original LBP in 6,76%. In relation to the binary sperm dataset, the best result was obtained with ALBPS and a kNN classifier (k=9), reaching a 72.66% of hit rate using the Chi Square metric, outperforming the original LBP in 22,47% and the ALBP in 1,22%. In the latter case, the weighted kNN did not improve the results achieved using just kNN. Taking this results into account, we can determine that ALBPS has more discriminant power for image retrieval than the rest of the tested LBP variants in different image contexts. © 2013 Springer-Verlag.