Self-similarity detection in noisy feature distribution

Abstract: A method is presented for detecting unknown fractal patterns in noisy imagery. Target patterns are assumed to be generated as the attractors of not-yet-identified contraction mappings. For such class of unknown patterns, invariant measure is introduced as a visualiza-tion of mapping structure. Based on the invariant measure, a smooth field, called the capturing probability, is defined within the framework of entropy maximization. Noise patterns added to target attractors are eliminated via two stages of image analysis: input- and output-filitering. The input-filter selects support points of unknown attractors via local complexity test. The output-filter is implemented as adaptive zero-crossing based on probabilistic complexity anal