Associative Learning of Standard Regularizing Operators in Early Vision

Standard regularization methods can be used to solve satisfactorily several problems in early vision, including edge detection, surface reconstruction, the computation of motion and the recovery of color. In this paper, we suggest (a) that quadratic variational principles corresponding to standard regularization methods are equivalent to a linear regularizing operator acting on the data and (b) that this operator can be synthesized through associative learning. The synthesis of the regularizing operator involves the computation of the pseudoinverse of the data. The pseudoinverse can be computed by iterative methods, that can be implemented in analog networks. Possible implications for biological visual systems are also discussed.MIT Artificial Intelligence Laborator