| Summary: | Sequential decision schemes for the purpose of both pattern classification and feature ordering are investigated.
An optimum compound sequential probability ratio test (OCSPRT) for recognition problems with memory in the source as well as the observation
medium is developed. The results of theoretical analysis and computer simulation of the OCSPRT for a two-class problem with first order Markov dependence among the pattern classes are presented.
For multiclass recognition problems the suitability of Bayes sequential decision schemes based on one-state ahead truncation approximation,
with and without on-line feature ordering, is assessed from the points of view of computational complexity and expected cost of a terminal decision. The Bayes sequential decision scheme for dependent hypothesis problems is formulated and its performance is compared with that of the optimum compound nonsequential decision scheme.
For dependent hypothesis recognition problems, compound sequential
pattern recognition schemes (CSPRS) are formulated. In CSPR schemes the required additional feature is observed either on the pattern to be decided, as in the classical sequential schemes, or on any one of the neighbouring patterns. The pattern selected and the feature actually observed are the ones which provide the maximum amount of additional information.
The results of analytical and experimental evaluation of the CSPR schemes are presented.
The suitability of the suboptimal sequential decision scheme with on-line ordering of features as a feature evaluation and ordering criterion is discussed. A modified on-line sequential (MOLS) decision scheme based on limited length of search is proposed as a compromise between the additional computational complexity and improvement in the recognition performance resulting from the on-line ordering of features. The advantage of incorporating such limited length of search over available features into sequential decision schemes using a set of preordered features is also examined.
For the purpose of experimental evaluation of the various decision
schemes, recognition of handprinted English text as a particular example of a pattern recognition problem was simulated on a digital computer. The handprinted characters were obtained from Munson's multiauthor data file prepared at Stanford Research Institute. === Applied Science, Faculty of === Electrical and Computer Engineering, Department of === Graduate
|
|---|
