Segment based stereo matching using cooperative Hopfield networks

This paper proposes a segment based dense stereo algorithm using cooperative Hopfield networks. It uses two Hopfield networks with similar structure to solve energy minimization problem in stereo matching parallely. In order to escape from local minima and speed the convergence of network, a Coarse-to-Fine strategy is employed. Firstly, the stereo image pairs are divided into non-overlapping homogeneous regions which can be represented as a set of layers in the disparity space. Then after each disparity layer is estimated, the more refined process is implemented in pixel domain. Experiments indicate this method has good performance and rapid convergence speed. Moreover it is insensitive to initial conditions of the neural networks and the neuron update orders.