Modulation domain image segmentation

Abstract

We present a modulation domain technique for segmenting textured images modeled as 2D multicomponent AM-FM functions. A Gabor filter bank is used to derive estimates of the dominant modulations at each image point and statistical clustering is then performed with respect to a similarity measure-based modulation domain entropy. Post processing is applied in the image domain to regularize the solution and enforce spatial correspondence. Several synthetic and natural examples are presented.