Image segmentation using fuzzy clustering with fractal features

Statistical features (mean and variance) are considered for region-based image segmentation. These features contain self similar structures which we interpret as fractal objects. Ellipsoidal attractors are transformed to approximately linear structures which are well separated and enable a robust segmentation. We identify class locations using fuzzy c-elliptotypes. The clustering results then yield segmentation using maximum membership defuzzification or, equivalently, a nearest prototype classifier. The method is applied to the digital mammograms from the Mammographic Image Analysis Society and produces reasonable segmentation in all cases.