Extracting essential modulated image structure

We present a new theory of multidimensional AM-FM image modeling and derive algorithms for extracting AM-FM sub-image information from digital images. In contrast to Fourier components, AM-FM image functions admit arbitrarily varying amplitude and phase modulations. Thus, they are inherently capable of efficiently capturing essential nonstationary image structures. Often, such nonstationarities contribute significantly to visual perception and interpretation. We describe a practical approach for computing AM-FM image representations using nonlinear demodulation operators. A Gabor filterbank isolates components locally, and optimal filters based on a statistical state-space component model are used to trade image multi-components across the filterbank channel responses. We present dramatic examples where the essential structure of natural images is successfully recovered from their computed AM-FM representations.