Robust Texture Image Representation by Scale Selective Local Binary Patterns.

Area of Science:

• Computer Vision
• Pattern Recognition
• Image Analysis

Background:

• Local Binary Pattern (LBP) is effective for texture recognition, handling grayscale and rotation variations.
• Traditional LBP methods struggle with scale variations in texture classification.
• Scale transformation poses a significant challenge in achieving robust texture recognition.

Purpose of the Study:

• To propose a novel method for texture classification that addresses scale variation.
• To enhance the scale-invariance of Local Binary Pattern (LBP) features.
• To achieve competitive accuracy in texture classification across diverse datasets.

Main Methods:

• A scale space of texture images is generated using a Gaussian filter.
• Histograms of pre-learned dominant Local Binary Patterns (LBPs) are constructed for each scale.
• Scale-invariant features are extracted by identifying the maximal frequency of dominant LBPs across scales.

Main Results:

• The proposed method achieved high accuracy on five public texture databases (UIUC, CUReT, KTH-TIPS, UMD, ALOT), with results ranging from 99.36% to 99.71%.
• The method demonstrates robustness against scale transformation, a known limitation of standard LBP.
• Scale-robust features are extracted efficiently, with a 200x200 image processed in 0.24 seconds.

Conclusions:

• The dominant LBP in scale space method significantly improves texture classification performance by addressing scale variation.
• The proposed approach maintains the efficiency and simplicity of LBP, making it suitable for real-time applications.
• This method offers a competitive and robust solution for scale-invariant texture recognition.