Radial mean local binary pattern for noisy texture classification

Local Binary Pattern (LBP) is one of the best descriptors of texture images; however, it elicits information from the pixels' value over each locality and therefore its value is highly sensitive to additive noise. In this research, a robust-to-noise LBP version is proposed, termed Radial Mean Local Binary Pattern (RMLBP), to enhance the quality of extracted features in noisy images. The main trick of RMLBP is to define the mean of points over each radial instead of using angular neighbor points (over a circle). This changing strategy enables RMLBP to extract robust features by removing the effect of noisy neighbors over each radial local patch. To make a fair comparison, the proposed method along with known mean filters, including circular and square mean, were applied to noisy textures. Applying RMLBP and the compared LBP variants to the Outex, CUReT and UIUC datasets demonstrated a significant superiority of the proposed method to its counterparts.