A compact information-theoretic framework for texture classification: Hilbert curves, amplitude-aware permutation...

Area of Science:

• Image Analysis
• Information Theory
• Machine Learning

Background:

• Texture classification is vital in remote sensing, materials science, and biomedical imaging.
• Characterizing spatial arrangements of pixel intensities is crucial for texture analysis.
• Existing methods may not fully capture amplitude information essential for discrimination.

Purpose of the Study:

• To propose a novel two-step information-theoretic framework for texture discrimination.
• To evaluate the effectiveness of various quantifiers, including amplitude-sensitive ones.
• To enhance texture classification accuracy by incorporating amplitude information.

Main Methods:

• Transforming images into 1D time series using space-filling Hilbert curves.
• Extracting eight complementary quantifiers: permutation entropy, statistical complexity, Fisher information, Wasserstein distances, weighted permutation entropy, and amplitude-aware permutation entropy.
• Training a support vector machine with nested cross-validation for hyperparameter optimization.

Main Results:

• The proposed framework successfully discriminates textures on the Kylberg database.
• While complexity-entropy planes show some discriminative power, amplitude-sensitive quantifiers are primary performance drivers.
• Weighted permutation entropy significantly improves texture classification accuracy.

Conclusions:

• Encoding amplitude information alongside ordinal patterns is essential for effective texture characterization.
• The developed framework offers a robust approach to texture discrimination.
• Amplitude-sensitive quantifiers represent a promising direction for future texture analysis research.