CHARACTER RECOGNITION SYSTEM

摘要

1437586 Character recognition systems INTERNATIONAL BUSINESS MACHINES CORP 12 July 1974 [25 Oct 1973] 30879/74 Heading G4R The system includes recognition units which respectively determine, for each scanned character, both the alphabetic character and the numeric character (here called the read characters) which most closely match the scanned character. The read characters address stores which contain tables of probability factors. The factors of one table indicate the probability of a read alphabetic character being a misrecognition of a scanned numeric character (which should appear as the read numeric character), and the factors of the other table the probability of a read numeric character being a misrecognition of a scanned alphabetic character. The factors read from the store are used to determine whether the read alphabetic or the read numeric character is more likely to be correct. The embodiment described examines groups of read characters, the groups forming for example the name and address on a postal item, and being separated by an interword space. The characters are scanned, normalized and segmented, and the lines in which the characters lie (e.g. the house number and road will be in line 2) are determined by known means, not described. Character features are extracted and passed to two recognition units which provide signals representing the read characters. If the number of rejects, i.e. characters which the recognition unit failed to identify, in the read alphabetic group is more than one greater than those in the corresponding read numeric group it is assumed that the read numeric group is correct, and vice versa. However if there is not a significant difference in the number of rejects the probability stores are both addressed by the sequence of read alphabetic and corresponding read numeric characters so as to produce, from each store, a sequence of probability factors. The factors of each sequence are multiplied together and the resulting two products compared to determine whether the alphabetic group of the numeric group is more likely to be correct and to gate the relevant group to a utilization unit. Before comparison takes place the probability products may be weighted by further factors which are selected from stored values in accordance with the line in which the group lies and in accordance with the position of the group within the line, e.g. the first group of the second line of an address is expected to be numeric.