Occluded object recognition by Hopfield networks

A new method to use a Hopfield neural network for object recognition is proposed. Object recognition is treated as a subgraph matching. A system consisting of one global network and several sub-networks is constructed. The sub-networks are dynamically changed and the outputs of the global network and the sub-networks are fedback to each other to complete the subgraph matching. This method avoids the local minimum problem arising from the use of one single Hopfield network and it also uses much less time than the simulated annealing algorithm. Computer simulation shows it can efficiently recognize objects in occlusion.<
>