cdohert@ccvax.ucd.ie (06/07/91)
Greetings fellow netters,
I am trying to write a report on the impact of neural network
research on OCR. I am familiar with a lot of the research papers
which report results along the lines of our "NN architecture
classified correctly 95% of our training/test set's experimenter
determined features" (oriented line segments, fourier descriptors
etc). The most promising results appear to come from the Bell Labs
group of LeCun et al. (1990) on zip code recognition using a local
receptive field backpropagation architecture to learn appropriate
classification features. I am not familiar with results from
classical pattern recognition using k-nearest neighbours or
whatever. Does anyone a) know of any papers evaluating Neural
Network approaches to OCR in the context of analytical pattern
classification's successes and failures if possible on common data
sets; b) know of any NN OCR systems which are in actual
commercial/military use? Are there any comments on what advantages
NNs bring to OCR apart from implementation (|| speedup) which I
imagine analytical techniques would also benefit from. Finally,
the key problem for correct classification of cursive script
appears to be segmentation. Is there any work on neural networks
performing segmentation?
Conor Doherty
Dept of Computer Science
University College Dublin
cdohert@ccvax.ucd.ie