Humans are well suited to the reading of textual information, but unfortunately it has not yet been possible to develop a machine to emulate this form of human behaviour. In the past, machines have been characterised by having static forms of specific knowledge necessary for character recognition. The resulting form of reading behaviour is most uncharacteristic of the way humans perceive textual information. The major problem with handprinted character recognition is the infinite variability in the character shapes and the ambiguities many of these shapes exhibit. Human perception of handprinted characters makes extensive use of "world knowledge" to remove such ambiguities. Humans are also continually modifying their world knowledge to further enhance their reading behaviour by acquiring new knowledge as they read. An information processing model for perception and learning of handprinted characters is proposed. The function of the model is to enable ambiguous character descriptions to converge to single character classifications. The accuracy of this convergence improves with reading experience on handprinted text. The model consists of three compon,ent parts. Firstly, a character classifier to recognise character patterns. These patterns may be both distorted anq noisy, where distortion is defined to be a consistent variability from known archetypical character descriptions and noise as a random inconsistent variability in character shape. Secondly, a perceptive mechanism that makes inferences from an incomplete linguistic world model of an author or of a specific domain of discourse from many authors. Finally, a incremental learning capability is integrated into the character classifier and perceptive mechanisms. This is to enable the internal world model to be continually adaptive to either changes in the domain of discourse or to different authors. A demonstrator is described, together with a summary of experimental results that clearly show the improvement in machine perception which results from continuous incremental learning.
|Creators||Malyan, R. R.|
|Source Sets||Ethos UK|
|Type||Electronic Thesis or Dissertation|
Page generated in 0.0141 seconds