Optical character recognition using artificial neural networks

Optical character recognition is examined to find a general framework by which it can be realized. A hierarchical 'cone' with feature extraction layers of increasing sophistication is described. The system, unlike the artificial neural net examples in the literature, does not use one network only. Allowing recognition to take place in parallel over different representations of the same symbol introduces redundancy, facilities learning and thus improves performance. The resource requirements of the system, which parallel operation inevitably increases, can be decreased by limiting the size of the image that can be 'seen' at one time. There is an 'eye' that can be moved around and fixed on any part of the scene which returns detailed information about a small part of the scene. The integration of successive eye fixations is a temporal process and the operation of the system also turns into that of relaxation in time where temporal expectations and selective attention should be taken into account. One possibility for representing spatial relations by introducing sequential scanning of the image is shown. Synapses with an internal delay, together with a temporal summation mechanism, are proposed by which this order can be checked. Work is currently going on to apply this mechanism to more realistic objects, feature sets, and scanning orders.<
>