Continuous Hopfield Neural Network Based Stereo Correspondence

A feasible approach to stereo correspondence based on continuous Hopfield neural network is proposed. It combines four constraints including similarity, uniqueness, ordering and smoothness in the proposed cost function in an energy form, which is mapped onto a continuous Hopfield neural network with appropriate interconnection weights between neurons. Furthermore, the minimization problem of the energy function can be converted into minimizing a cost function representing the dynamics of the network. The minimization is obtained when the dynamic system is at its stable state. The experimental results have shown its feasibility and effectiveness, where the proposal is compared with dynamic programming method for the very similar constraints used in both of these algorithms.