Robust Adaptive Median Binary Pattern for Noisy Texture Classification and Retrieval.

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Texture analysis is crucial for computer vision.
• Local Binary Pattern (LBP) is an efficient texture descriptor but is sensitive to noise.
• Existing methods struggle with noisy texture data.

Purpose of the Study:

• Introduce a novel texture descriptor, Robust Adaptive Median Binary Pattern (RAMBP).
• Address the limitations of existing texture descriptors, particularly noise sensitivity.
• Enhance texture classification and retrieval performance on noisy images.

Main Methods:

• Developed RAMBP using noisy pixel classification, adaptive analysis windows, scale analysis, and median comparison.
• Evaluated RAMBP on standard texture datasets under various noise conditions (impulse, Gaussian, blur, mixed).
• Assessed performance without prior noise type knowledge or training.

Main Results:

• RAMBP achieved superior classification accuracy (>90%) even with 50% impulse noise.
• Demonstrated high performance (>95%) on Gaussian and blurred textures.
• Outperformed state-of-the-art methods in noisy texture classification and retrieval.
• Exhibited competitive running time and low feature dimensionality.

Conclusions:

• RAMBP is a robust and effective texture descriptor for noisy images.
• The method offers improved discriminative properties by capturing multi-scale texture information.
• RAMBP presents a significant advancement for texture analysis in challenging, real-world conditions.