Bayesian segmentation of AM-FM texture images

We present a fully unsupervised parametric modulation domain technique for segmenting textured images. Textured regions in the image are modeled as multicomponent sums of nonstationary AM-FM functions. The dominant modulations at each pixel are estimated using a technique called DCA and used to construct modulation domain feature vectors. The overall feature space is regarded as a mixture of Gaussians, where the modulations within each texture class are modeled by a single multivariate normal distribution. Although this model is somewhat unrealistic, it leads to a robust segmentation algorithm that is able to operate in a fully unsupervised mode. An EM algorithm is used to estimate the parameters of the Gaussian mixture so that approximate maximum-likelihood estimates of the pixel class labels can be obtained. The proposed technique is demonstrated on a variety of images constructed from juxtapositions of Brodatz-like textures.