Extended fractal analysis for texture classification and segmentation

Area of Science:

• Image analysis and computer vision
• Statistical modeling of natural phenomena

Background:

• Fractional Brownian motion (fBm) uses a single Hurst parameter for scale-invariant textures.
• Extended self-similar (ESS) processes generalize fBm with multiscale Hurst parameters for non-scale-invariant textures.

Purpose of the Study:

• Evaluate multiscale Hurst parameters for texture classification and segmentation.
• Compare generalized Hurst features against traditional Hurst and Gabor features.
• Assess segmentation performance on texture mosaics and real synthetic aperture radar (SAR) imagery.

Main Methods:

• Texture classification using generalized Hurst, standard Hurst, and Gabor features.
• Texture segmentation on mosaic images with and without supplemental contrast/grayscale features.
• Application of Hurst features for segmenting real SAR imagery.

Main Results:

• Generalized Hurst features achieve comparable classification accuracy to Gabor features at a fifth of the dimensionality.
• Segmentation accuracy is evaluated for generalized and standard Hurst features on texture mosaics.
• Hurst features demonstrate effectiveness in segmenting real SAR imagery.

Conclusions:

• Multiscale Hurst parameters are effective features for texture classification and segmentation.
• Generalized Hurst features offer a dimensionality reduction advantage.
• These parameters show utility in analyzing diverse natural textures and specialized imagery like SAR.