Training-Based Gradient LBP Feature Models for Multiresolution Texture Classification.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Local Binary Pattern (LBP) is a foundational technique for texture feature extraction.
• Existing LBP methods face limitations in classification accuracy and noise robustness.

Purpose of the Study:

• To enhance texture classification by developing novel gradient Local Binary Pattern (gLBP) descriptors.
• To introduce a new feature mapping method and a multiresolution feature fusion framework.

Main Methods:

• Designed a median sampling regulation and defined gradient LBP (gLBP) descriptors (central, radial, magnitude, tangent gradients).
• Proposed a training-based feature model mapping using maximal relative-variation rate (mr2) for dimensionality reduction.
• Developed a multiresolution feature fusion approach using concatenated gLBP descriptors.
• Employed a nearest neighbor classifier with chi-square distance for texture classification.

Main Results:

• The proposed gLBP method demonstrated superior performance across five public texture databases.
• Achieved higher classification accuracy and improved robustness against Salt&Pepper and Gaussian noise compared to nine other methods.
• Outperformed state-of-the-art approaches in both accuracy and noise resilience.

Conclusions:

• The developed gLBP-based texture classification framework is reliable and efficient.
• The method offers significant advantages in terms of speed, accuracy, and noise robustness.
• This approach represents a notable advancement in texture analysis and classification.