Solving the correspondence problem using a Hopfield network

This paper proposes a Hopfield network to solve the correspondence problem in computer vision. The correspondence problem is that of identifying features in two images that are projections of the same entity in the 3D world. In this paper, correspondence is established between the edge points in the two images. The problem is first formulated as an optimization problem where an energy function is to be minimized. The optimization problem is then solved using a Hopfield network. The technique can also be considered as a constraint satisfaction process, where the nodes are the hypotheses (possible correspondences), and the links between them are the constraints. The weights of these links are derived by making some assumptions about the objects being imaged. Once started, the network converges rapidly to a good solution. The approach presented here is much simpler, and more elegant compared to the other techniques proposed in the literature. It is also easier to design and implement, and is more suitable for parallel implementation. The efficacy of this approach is demonstrated with experimental results on stereo images