A multiscale approach to sensor fusion and the solution of linear inverse problems (Q1893657)

An approach to the solution of linear inverse problems is presented. The authors develop a framework for inversion based upon a multiscale and stochastic description of the data, the operators, and the function to be reconstructed. It is assumed that the data upon which the inversion is to be based, \(y_i (x)\), are related to the function to be reconstructed, \(g(x)\), via a system of linear integral equations embedded in additive noise. The observation model is \(y_i (x) = \int T_i (x,x') g(x') dx' + n_i (x)\), \(i = 1,2,\dots,K\), where the integral kernels, \(T_i (x,x')\), and the characteristics of the noise processes, \(n_i(x)\), are known. The inversion algorithm is drawn from the theory of statistical estimation. This approach allows for the explicit modeling of the errors in the data as simple paths from random processes. All prior information regarding the structure of the underlying function is summarized in the form of a statistical model which also acts as a regularizer. The authors introduce the relative error covariance matrix. This matrix is central to an understanding of the manner in which information from a set of observations is propagated into a reconstruction. A \(1/f\) fractal prior model is specified in the wavelet transform domain for the purpose of regularization. The use of a scale space approach is suited to the fusion of data from several sensors with differing resolutions and spacial coverage; so the authors develop a quantitative theory of sensor fusion. The paper of \textit{G. Beylkin, R. Coifman} and \textit{V. Rokhlin} [Fast wavelet transforms and numerical algorithms. I, Comm. Pure Appl. Math. 44, No. 2, 141-183 (1991; Zbl 0722.65022)] and their results on the compression of whole classes of linear operators in a nonstandard wavelet representation are essentially used. Results of computer experiments as well as a brief survey of the subject are presented. The list of references contains 60 items.