Local Directional Pattern (LDP) – A Robust Image Descriptor for Object Recognition

Abstract

This paper presents a novel local feature descriptor, theLocal Directional Pattern (LDP), for describing localimage feature. A LDP feature is obtained by computing theedge response values in all eight directions at each pixelposition and generating a code from the relative strengthmagnitude. Each bit of code sequence is determined byconsidering a local neighborhood hence becomes robust innoisy situation. A rotation invariant LDP code is alsointroduced which uses the direction of the most prominentedge response. Finally an image descriptor is formed todescribe the image (or image region) by accumulating theoccurrence of LDP feature over the whole input image (orimage region). Experimental results on the Brodatz texturedatabase show that LDP impressively outperforms theother commonly used dense descriptors (e.g.,Gabor-wavelet and LBP).