Median Robust Extended Local Binary Pattern for Texture Classification

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Local Binary Patterns (LBP) are efficient texture features but sensitive to noise and lack macrostructure detail.
• Existing LBP variants struggle with noise robustness and capturing broader texture patterns.
• Limitations in traditional LBP hinder its application in challenging image conditions.

Purpose of the Study:

• To introduce a novel texture descriptor, Median Robust Extended LBP (MRELBP).
• To address the noise sensitivity and macrostructure limitations of traditional LBP.
• To develop a computationally efficient and robust texture classification method.

Main Methods:

• Developed MRELBP by comparing regional image medians instead of raw intensities.
• Implemented a multiscale descriptor using a novel sampling scheme for capturing diverse texture information.
• Evaluated MRELBP on benchmark datasets including Outex test suites.

Main Results:

• MRELBP demonstrated high robustness to grayscale variations, rotation, and various noise types (Gaussian, blur, salt-and-pepper, random corruption).
• Achieved excellent classification scores: 99.82% (Outex-1), 99.38% (Outex-3), and 99.77% (Outex-4).
• Maintained high performance at a low computational cost compared to traditional LBP.

Conclusions:

• MRELBP effectively overcomes the noise sensitivity and macrostructure limitations of traditional LBP.
• The proposed descriptor offers superior texture classification performance and robustness.
• MRELBP is a promising feature descriptor for real-world applications with noisy or varied image data.