Application of Multifractal analysis for image analysis
CREATED: 201003301039
Statistical self similarity: having certain statistical properties or mesasures (Hurst parameter, roughness, etc) that are preserved across various scales
Fractal dimension, D: the number of features of a certain size l varies as l^-D For a line, D = 1
Local density: eg sum pixel intensity around a point p with radius r, u_r(p) proportional to r^a. a is the Lipschitz-Holder exponent. For a mono-fractal, the holder exponent is constant.
Multi-fractal: collection of several fractal structures with varying dimensions.
Biomedical images: non-linear, non-stationary and noisy, contains highly irregular structures, features must represent granularity and regularity of the structure
Computing coarse holder exponent:
- compute intensity based measure defined inside a square window of size r, centered at p
- do a linear regression on log(r), log(u) to determine a
Notion of alfa image (replace intensity at point p with value of alfa at p)
Compute fractal dimension using box counting method.
Multi-fractal spectrum is the relationship between a and corresponding D.
Applications
- difference in multi=fractal spectrum between diseased and normal retina tissue.
- edge detection
- DNA sequences
- sleep EEG signals
- segmentation of digital mammograms
Tissue image classification, retrieval
- alfa histogram as a image feature
- reduced alfa histogram (keep only 5 values)
- multi-fractal spectrum
- approximate multi-fractal spectrum using piecewise cubic polynomials, changes the spectrum to a continuous curve and make it easier to compare two spectrums without interpolation
Multi-fractal measures
- sum of intensities
- normalized sum of intensities
- iso measure
- maximum intensity
- inverse minimum
- sum of absolute difference
HRCT image segmentation using alpha slices (set of pixels within a certain range of alpha values)