Pairwise Rotation Invariant Co-Occurrence Local Binary Pattern.

Area of Science:

• Computer Vision
• Image Analysis
• Machine Learning

Background:

• Designing effective features is crucial in computer vision, but achieving a balance between discriminative power and robustness is challenging.
• Existing co-occurrence features enhance discriminative power but lack robustness against geometric and photometric variations.

Purpose of the Study:

• To investigate Transform Invariance (TI) in co-occurrence features.
• To propose a novel feature, Pairwise Rotation Invariant Co-occurrence Local Binary Pattern (PRICoLBP), that offers improved robustness and discriminative power.

Main Methods:

• Formal introduction of the Pairwise Transform Invariance (PTI) principle.
• Development of the PRICoLBP feature, incorporating multi-scale, multi-orientation, and multi-channel information.
• Comprehensive evaluation on nine benchmark datasets across various perspectives including encoding strategy, rotation invariance, speed, and discriminative power.

Main Results:

• PRICoLBP effectively captures spatial context co-occurrence information while maintaining rotation invariance.
• The proposed feature demonstrates a well-balanced tradeoff between discriminative power and robustness.
• Evaluations show PRICoLBP outperforms other Local Binary Pattern (LBP) variants in various classification tasks.

Conclusions:

• PRICoLBP offers an efficient and effective solution for feature design in computer vision.
• The feature's robustness and discriminative capabilities make it suitable for diverse applications like texture, material, and scene classification.