Noise robust and rotation invariant texture classification based on local distribution transform

Applying local binary pattern (LBP) to images with uniform distribution leads to generate discriminative features; however, the distribution of all images is not necessarily uniform. The distribution of an image can be uniformzed if it passes through its cumulative distribution function (CDF), while estimation of CDF is highly sensitive to additive noises. In this paper, we propose a novel transform, which locally uniformize all patches of an image and approximately estimate a robust CDF. The proposed local distribution transform (LDT) generates continuous values and by quantizing them into discrete values, a histogram of features is constructed. We have fused the LDT features to the features of rotation invariant LBP and local variance (VAR) in order to provide a rich set of robust-to-noise features, which can detect both uniform and non-uniform patterns. The performance of the proposed LDT-LBP_VAR is assessed over a wide range of datasets like Outex, UIUC, CUReT, Coral Reef, Virus and ORL. The datasets are also corrupted by additive Gaussian noise with different signal to noise ratio (SNR) and the empirical results demonstrate that the proposed hybrid features provide superior classification results (P < 0.05) to the plenty of advanced descriptors over the datasets in both noise-free and noisy conditions.