Noise-tolerant texture feature extraction through directional thresholded local binary pattern

Abstract

Local binary pattern (LBP) is a multi-applicable texture descriptor applied in machine vision. Despite its outstanding abilities in revealing textural properties of image, it is sensitive to noise, due to its thresholding mechanism. To make LBP robust against noise, a directional thresholded LBP (DTLBP) is developed in this article which applies the directional neighboring pixels average values for thresholding. Applying this type of thresholding in addition to reducing noise, due to using the information of neighboring pixels with bigger radii, increases efficiency in extracting features. The DTLBP is able to be combined with other descriptors like completed LBP (CLBP) and local ternary pattern (LTP) which improves their functionality against noise. To evaluate the functionality of DTLBP, four known datasets including Outex (TC10), CUReT, UIUC and UMD are tested. Numerous and extensive experiments on these datasets with different kinds of noises indicate this newly developed descriptor’s efficiency, with or without incremental white Gaussian and Gaussian blur noises. The proposed descriptor is compared with its available state of the art counterparts. The results show that the combination of DTLBP with CLBP descriptors provide the best classification accuracy in the experiments, which confirms the efficiency and robustness of the proposed descriptor when extracting features from noisy and raw images.