This paper is concerned with the recognition of cursive handwritten words selected from words in a dictionary and written off line. This problem arises in post office applications when addresses must be read from letters and packages. Off-line recognition is a much more difficult problem than the recognition of on-line cursive handwriting, where the signal can be considered as one-dimensional.

The off-line cursive script problem is interesting because it combines elements of visual perception with those of language perception and under- standing. The authors approach the problem as one of multi-level perception in which a cursive script word image is transformed through a representa- tional hierarchy. The recognizer of off-line cursive script (ROCS) processes the input hierarchically in a bottom-up fashion until a certain level is reached, and processing is heterarchical after that, with information flowing in both directions.

On the image level, ROCS first removes the slant of strokes and then removes minor contour discontinuities and roughness. Three main vertical zones of script are found and the word is partitioned horizontally into minimal portions that are potential letters or parts of letters, each called a presegment. A character can span one or more presegments. The contour level of analysis describes the different contours of the word and their global topology; an entry contains its topological depth level and its chain encoding. Event level analysis determines 16 features and their locations in the word in a single scan of the contour level encodings. The event-level representation is a binary array whose rows are overlapping presegments and whose columns represent features. Thus, sequential rows of this array might represent presegments 0–1, then 0–2, then 1–2, and each column would have a 1 or a 0 to indicate the presence or absence of a feature.

In the final step, the event-level array is examined from top to bottom, and letter hypotheses (rated on a −1 to +1 scale) are computed. Hypothesizing a letter depends on program parameters that are determined during training. The parameters define the probabilistic rules of letter formation from the primitives of the event set. The hypotheses are sent to the lexicon lookup where inadmissable hypotheses are pruned, and the one remaining with the highest rating is expanded. The process proceeds iteratively and candidate letter sequences are revised until a match with a lexicon item is obtained that accounts for the entire word (all presegments).

The authors throw an impressive amount of technology at the problem, but the final results are nothing to write home about. In tests on a lexicon of 710 words where the same writer is used for both training and testing, correct recognition is about 77 percent, with 8 percent incorrect, and 14 percent rejected. The performance of the system is affected by the size of the lexicon; for a 7,800 word lexicon the same test situation resulted in 48 percent correct, 31 percent incorrect, and 21 percent rejected. When the system is trained on one writer and tested using another writer, the system has about 30 percent correct recognition, 38 percent incorrect, and 33 percent rejected. The system performance improves 50–60 percent when retrained with the new writer, and the training process is not as lengthy.

The problem of recognizing off-line cursive script must indeed be a difficult one to withstand the ROCS armamentarium] This long paper clearly describes the many techniques used in ROCS, and a complete set of references is given.