High-performance OCR preclassification trees

We prese nt an automatic method for construc ting high-per forma nce pre classifica tion decision trees for OCR. Good pre classifier s must prune the set of alter native classes to a small number without err oneously pruning the corr ect class. We build the decision tree using gree dy entropy minimization, using pseudo-r andomly gener ated training samples der ived from a model of imaging defe cts, and then ‘‘ populate’’ the tree with many more samples to drive down the er ror ra te. We descr ibe a re fineme nt of the method of [BM94] that approa ches the userspecifie d acc urac y more closely and thus allows higher pruning. The essential technica l device is a leaf -sele ction rule based on the Good-Tur ing Theore m [Good53]. Such a prec lassifier , constructe d for a panEuropea n polyfont classifie r, attains a 1% err or ra te and a 3.8 pruning fac tor, in tests on synthetic images. On re al pages printed in ten Europea n languages, the pre classifier sped up the page rea der by a fa ctor of 2.2, with no mea surable incre ase in er ror.