Degraded Printed Word Recognition with a Hop eld-style Network

We have previously developed and analysed a new Hop eld-style network model. This model has the advantage of higher storage capacity and less interference between stored memories than the standard discrete Hop eld network. In this paper, this model is applied towards machine printed word recognition. Words to be recognised are stored as content-addressable memories. Word images are rst processed by an OCR. The network is then used to postprocess the OCR decisions. It is shown that when the number of stored words is small, the network can reliably recall the correct word and when the number of stored words is large, exact recall performance deteriorates but the network can still be used to postprocess the OCR output with very good performance. Lastly, very large dictionaries are handled by partitioning them by word length and storing them in multiple networks and this scheme gives very good performance.