Generalization of the Harris `coupled depth-slope´ analog visual reconstruction network

A powerful computational paradigm in computer vision research is that one should formulate the various reconstruction problems as the minimization of a functional that characterizes the degree of acceptability of a solution according to the existing constraints. The analogy between the energy of an analog network and the value of the functional to be minimized leads to natural analog neural network implementations. J.G. Harris (1987) provided an analog network that contained layers corresponding to the case where more than one derivative is included in the smoothness term. However, his mathematical justification relied on a simple penalty-based approach to ensure compatibility between the derivatives. The author shows how a more general approach based on augmented Lagrangian formulations can be used to derive similar networks including that of Harris as a special case. The Harris coupled depth-slope analog model of visual reconstruction is discussed