Hyperspectral Texture Metrology Based on Distance Measures in an Information-Theoretic Framework.

Area of Science:

• Image analysis and machine learning
• Spectroscopy and optical sensing
• Computational physics

Background:

• Existing hyperspectral texture feature extraction methods often require pre- or post-processing.
• Dimensionality reduction techniques like PCA, ICA, and NMF can lead to information loss.

Purpose of the Study:

• To integrate metrology into hyperspectral texture feature extraction.
• To propose novel, direct, and interpretable feature calculation methods for hyperspectral images.
• To introduce a new spectral distance metric, Radical of Extended Mean Information for Discrimination (REID).

Main Methods:

• Developed distance-based expressions for Gray-Level Co-occurrence Matrix (GLCM), Local Binary Pattern (LBP), and Gabor filtering.
• Introduced REID, a novel spectral distance metric decomposable into spectral direction and magnitude.
• Feature calculations are fullband, lightweight, and interpretable, utilizing all wavelengths.

Main Results:

• The proposed methods consistently outperform existing techniques in texture classification, content-based image retrieval, and cancer detection.
• Achieved superior performance compared to methods using dimensionally reduced spaces (PCA, ICA, NMF).
• Demonstrated efficiency and interpretability of fullband feature calculations.

Conclusions:

• The novel hyperspectral texture feature extraction methods offer significant improvements over existing approaches.
• The REID metric provides a robust and interpretable way to analyze spectral information.
• These methods have potential for integration into machine/deep learning systems for explainable AI (XAI).