SOME STATISTICAL ASPECTS OF CHARACTER RECOGNITION.

The character recognition problem is considered as a generalization of the discrimination problem. Two issues that are of trivial importance in the two category problem become central to the n-category problem. Not all of the available information is relevant to each of the discriminations that have to be made. A few of the discriminations are much more difficult that all of the others. The analysis consists mainly of studying which character pairs are difficult to separate. After dealing with these it is relatively easy and expedient to separate the others by a few functions of the observations which are less than optimal for a particular pairwise discrimination but which serve to make many pairwise discriminations adequately. Assuming a particular physical mechanism and a specific character set of 36 characters, decisions are made on the choice of blocksize in the scanning area, the procedure for positioning characters within the scanning area, the choice of twelve linear functions of the blackness in a block for each of the blocks in terms of which the discriminations are to be made, and the robustness of the system to variations in character position. Author