2D-LCoLBP: A Learning Two-Dimensional Co-Occurrence Local Binary Pattern for Image Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Feature invariance (rotation, scale, translation) is crucial for image recognition.
• Local Binary Patterns (LBP) are efficient but often lack scale invariance and struggle with feature discrimination vs. dimension trade-offs.

Purpose of the Study:

• To propose a novel learning 2D co-occurrence LBP (2D-LCoLBP) descriptor.
• To address limitations of existing LBP methods, specifically achieving scale invariance and a better trade-off between feature discrimination and dimension.

Main Methods:

• Constructing a weighted joint histogram for multi-neighborhood and multi-scale LBP (2D-MLBP) to ensure rotation invariance.
• Employing a feature learning strategy on LBP pattern pairs across scales to achieve scale invariance, reduce dimensions, and enhance robustness.
• Utilizing a linear Support Vector Machine (SVM) classifier for recognition.

Main Results:

• The proposed 2D-LCoLBP descriptor achieves rotation and scale invariance.
• Demonstrates significantly lower feature dimensions compared to state-of-the-art LBP descriptors.
• Outperforms existing LBP-based methods in accuracy across various image recognition tasks (texture, object, face, food) under different noise conditions.

Conclusions:

• 2D-LCoLBP offers a superior approach to feature extraction for image recognition.
• The method effectively balances feature discrimination, dimension reduction, and robustness to scale variations and noise.
• Achieves state-of-the-art performance with reduced computational complexity.