Computing multi-purpose image-based descriptors for object detection: Powerfulness of LBP and its variants

Abstract

In this paper, we present some powerful methods for computing multi-purpose image-based descriptors toward their exploitations in object detection and recognition applications. Image-based descriptors characterize image properties for constituting machine learning-based recognition systems. In this context, we present the principle for computing image-based descriptors using the local binary pattern (LBP) method. Such a method is multi-purpose in the sense that it can be efficiently exploited for the automated recognition of objects of varied natures (e.g., vehicles of traffic-monitoring images, cells of medical images). Then, we propose three variants of LBP, named Mean-LBP, $$\lambda $$ -2RLBP, and C2R-LBP. The two latter ones use Hamming distance. Experimental results show that our method can overpass performances of discussed LBP methods notably under realistic conditions of use.