Methods for Updating of Model Parameters Applied within the Area of Symbol Recognition

Use of statistical classi cation methods in symbol recognition requires the speci cation of a number of model parameters Such parameters are usually ob tained from a training set This paper describes meth ods for improving parameter estimates obtained from a training set The improvement is made by us ing new observations for updating the parameter es timates The class membership of these new observa tions may either be known or unknown The potential of parameter updating techniques within the area of symbol recognition is discussed and tested on some examples The results of using parame ter updating for automatic training are promising and may lead to a considerable simpli cation of the train ing process The problem of adapting the parameter estimates to a slightly changed situation like for in stance new fonts did not lead to that good results Nevertheless the results showed that also here there is a potential if the methods are further developed