• Proposes EMSABP: a unified multi-scale adaptive binary pattern descriptor. • Uses Gaussian-weighted interpolation for improved noise robustness. • Employs adaptive sign, magnitude, and center encoding for stability. • Achieves consistently strong classification accuracy on Outex, UIUC, XUHR, UMD, and CUReT benchmarks. • Demonstrates competitive performance against advanced handcrafted descriptors and selected deep learning baselines. Texture classification is a fundamental problem in computer vision with applications in medical imaging, remote sensing, and industrial inspection. Local Binary Pattern (LBP) descriptors are widely used due to their simplicity and discriminative power, but they remain highly sensitive to noise, rely on fixed thresholds, and struggle with complex or multi-scale textures. This paper proposes the Enhanced Multi-Scale Adaptive Binary Pattern (EMSABP), a unified texture descriptor that extends LBP through three key improvements: (i) Gaussian-weighted interpolation to reduce noise sensitivity, (ii) local mean-based sign encoding for stable binary representation, and (iii) adaptive variance-driven magnitude thresholding to capture both weak and strong contrast variations. In addition, EMSABP integrates sign, magnitude, and center components within a multi-scale framework that concatenates power-normalized histograms across multiple radii, enabling the joint modeling of fine- and coarse-scale structures. The novelty of EMSABP lies in its unified integration of these techniques into a single handcrafted descriptor, simultaneously improving noise resistance, scale adaptability, and discriminability. Experiments on benchmark datasets (Outex (TC10, TC12, TC13), UIUC, XUHR, CUReT and UMD) demonstrate that EMSABP achieves higher accuracy than advanced handcrafted descriptors and remains competitive with selected deep learning baselines, while maintaining moderate feature dimensionality and efficiency.
Gupta et al. (Tue,) studied this question.