Robust color texture descriptor for material recognition

Abstract

This paper addresses the task of material and natural texture classification. We propose a new discriminant color texture descriptor based on local pattern encoding scheme using local maximum sum and difference histograms. This descriptor aims to improve the robustness of the sum and difference histograms (SDH) operator to noise and low quality images by using adaptive values of the pixels instead of their exact values. We also enhance the time and memory efficiency as well as the robustness to scale and rotation variations by rectifying the formula used to compute the SDH. Finally, we use three different encoding schemes in order to incorporate both color and texture information into the final descriptor. The proposed approach is evaluated on six popular and challenging datasets: Outex-TC13, New-BarkTex, FMD, KTH-TIPS2b, Outex-TC30 and Outex-TC32 databases. The experimental results show that the introduced operator significantly outperforms the state-of-the-art methods and enhance the performance of the basic SDH.