Classification of Textures Using a New Descriptor Circular and Elliptical-LBP (CE-ELBP)

摘要

The Local binary Pattern (LBP) is a very simple and popular approach, and it played a vital role in many image processing applications. LBP captures isotropic structural information. The LBP completely fails in representing anisotropic information. Later, to represent anisotropic, horizontal elliptical LBP (H-ELBP) and vertical elliptical LBP (V-ELBP) are derived, however they derived only partial anisotropic information. To derive complete anisotropic information these two ELBPs are concatenated and it increases the feature vector size by two folds when compared to LBP. To capture both isotropic and anisotropic structural information, one needs to concatenate the histograms of LBP, H-ELBP and V-ELBP and this process increases the feature vector size into three folds when compared with LBP. We present a novel Feature Extraction method known as "circular and elliptical-LBP (CE-ELBP)" and a new variant of local binary pattern (LBP) and elliptical LBP (ELBP). The CE-LBP captures both isotropic and anisotropic structural information with a feature vector size equivalent to LBP. The CE-LBP quantizes/places the multi-structure information of LBP and ELBP (H-ELBP and V-ELBP) and derives a unique CE-LBP code that represents the complete set of micro patterns. The distinct advantages of CE-LBP are its ease in implementation, invariance to monotonic illumination changes and low computational complexity. The CE-LBP is tested on well-known databases i.e., Brodtaz, UIUC and Outex using machine learning classifiers. The classification results are compared with conventional LBP, H-ELBP, V-ELBP and combination of these descriptors. The uniform patterns are also derived on CE-LBP and classification results are derived and compared. The results indicate the efficacy of the proposed method.