Bayesian unsupervised multifractal image segmentation using a multiscale graph label prior

This paper presents a Bayesian multifractal segmentation method that segments multifractal textures in regions with different multifractal properties. First, a computationally and statistically efficient model for wavelet leader-based multifractal parameter estimation is developed, assigning wavelet leader coefficients associated with distinct parameters to different image regions. Next, a multiscale graph label prior is introduced to capture spatial and scale correlations among these labels. Gibbs sampling is used to generate samples from the posterior distribution. Numerical experiments on synthetic multifractal images demonstrate the effectiveness of the proposed method, outperforming traditional unsupervised and modern deep learning-based segmentation approaches.