Multi-dimensional multivariate Gaussian Markov random fields with application to image processing (Q1098490)

The author investigates properties of multivariate Gaussian Markov random fields (GMRF) on multidimensional lattices. First he relates the distribution of conditional normal variates \(x_ i| \{x_ j\); \(j\neq i\}\) \((i=1,...,n)\) with the joint distribution of \(x'=(x'_ 1,...,x'_ n)\) in terms of the coefficients \(\beta_ i=(\beta_{ij})\) in \[ E[x_ i| x_ j;\quad j\neq i]=\mu_ i+\sum^{n}_{j\neq i}\beta_{ij}(x_ j-\mu_ j) \] and of \(\Gamma_ i=Var(x_ i| x_ j\); \(j\neq i)\). For Markovian fields, conditioning with respect to \(\{x_ j\); \(j\neq i\}\) reduces to conditioning with respect to \(\{x_ j\); j neighbour of \(i\}\). The author gives in particular characterizations of stationary GMRF by means of the spectral density matrix and stationarity conditions for specific GMRF. The last section is devoted to the estimation of \(\beta\) and \(\Gamma\) for an image processing example.