A multi-scale approach for early vision and its implementation by neural networks

Physics-based deformable model is a general framework for the ill-posed problems in early vision. In these deformable models based methods smoothness constraints can be implicitly controlled by a scale parameter. We present here a new multiscale method where scale is both time and space variant, thus it is rarely trapped at local energy minima and can significantly preserve discontinuities in the solution We also introduce a new parameter that can make external energy vanished at some space points where measured data seem to be very unreliable. Finally we propose a greedy algorithm to solve the minimization of energy functional of deformable model, and its implementation by neural networks. We applied our method to data regularization and blind deconvolution problems with encouraging results