Performance analysis of cooperative hopfield networks for stereo matching

Abstract

This paper proposes a dense stereo matching algorithm based on cooperative Hopfield networks. It uses two Hopfield networks with similar structure to solve energy minimization problem of stereo matching in parallel. Two strategies are applied to the performance analysis. One strategy considers each pixel as a neuron. The other is the Coarse-to-Fine strategy, which firstly divides the images into non-overlapping homogeneous regions, and each region is represented as super-pixel of the coarse images. After coarse estimation, a more refined estimation is implemented in pixel domain. Experiments indicate the method with the Coarse-to-Fine strategy has better performance and more rapid convergence speed, and less insensitive to initial conditions of the neural networks and the neuron update orders. © Springer-Verlag Berlin Heidelberg 2007.