Allen R. Hanson

Context in word recognition

Abstract Relatively low character error rates can often lead to prohibitive levels of word error rates. This paper examines several techniques for integrating an independent contextual postprocessor (CPP) into a full classification system. Using positional binary n -grams the CPP can correct many errors directly. In those cases where the correction process leads to ambiguity, the CPP can direct additional processing. Experimental results demonstrate that almost all of the derived improvement results from CPP-directed reclassification. This only requires that the CPP have the classifier likelihood fed forward to it. Therefore, a standardized CPP can be built independently of the rest of the classification system. An initial 45% word error rate is reduced to about a 2% word error rate and a 1% reject rate. Presence of a dictionary allows these figures to be reduced even further.

Variations in relaxation labeling techniques

Several issues concerning the behavior of relaxation labeling processes are examined in this paper. The performance of region relaxation is shown to be affected by the form of compatibility coefficients and the configuration of local neighborhoods. Three variants of the algorithm are explored on an artificially generated test image: relaxation with simple unary compatibility coefficients, probabilistic relaxation with piecewise-linear compatibility coefficients derived from conditional probabilities, and a form of discrete relaxation without compatibility coefficients. The compatibilities based upon conditional probabilities yield the only version that preserves image detail associated with spatially thin structures.