Resolution-invariant separable ARMA modeling of images

Abstract

We suggest a continuous-domain stochastic modeling of images that is invariant to spatial resolution. Specifically, we are proposing an estimator that is calibrated with respect to the sampling step, and that can potentially handle aliased data. Motivated by Markov random fields, we assume a continuous-domain ARMA model and suggest an algorithm for estimating the continuous-domain parameters from the sampled data. The continuous-domain parameters we estimate provide features that can further be used for image classification, segmentation and interpolation, regardless of sampling interval values and of aliasing effects that may appear in the digital image. Experimental results indicate that the proposed approach is preferable over a discrete-domain ARMA modeling.

Linear interpolation of biomedical images using a data-adaptive kernel

“…These parameters can then be used in (7). We rely on results from [9] that describe 2D separable AR models, as given next. The autocorrelation sequence of the sampled model is given in the z-domain by The green squares, the red circles, and the blue dots correspond to the best Gaussian-likelihood, the best l 2 -autocorrelation, and the best oracle-SNR parameter fits for our continuous-domain AR model, respectively.…”