TLBP: Tomography-Aided Local Binary Patterns With High Discrimination for Image Classification

Abstract

Local binary patterns (LBP) play a vital role in image classification as a computationally efficient feature descriptor. A crucial reason for its limitation of discriminability is the lack of neighbourhood information description from a global perspective. Previous research has attempted to improve its performance by introducing global thresholds, but such threshold selection is not optimal. To address this issue, we propose a novel tomography-aided local binary patterns (TLBP), inspired by the tomographic process of sample separation. TLBP considers constructing visual feature representations under multi-level non-local information to compensate for the lack of LBP possessing only a single shallow feature. In addition to the basic LBP features from local visual context, TLBP captures refined neighbourhood greyscale information through multi-quantile thresholds from a global visual perspective, thereby greatly enhancing discriminability. Experimental results in texture classification, face recognition, and hyperspectral pixel-wise classification demonstrate that the proposed TLBP descriptor outperforms the competitors, achieving 94.39% (KTH-TIPS), 81.22% (KTH-TIPS-ROT), 93.81% (Indian Pines), 99.85% (Salinas), and 99.50% (ORL) accuracy. Furthermore, the performance of the T-variants that apply the tomographic idea to classic LBP descriptors improve significantly, especially for their rotation-invariant versions.