The title of this paper is slightly misleading, especially since one of the keywords listed is topological features. The phrase, “simply connected,” does not refer to the topology of a handwritten numeral, but to the connection stroke between numerals, such as in a string of 0s. It is the connection by a single stroke rather than by, perhaps, characters inadvertently touching at more than one point. In fact, an attempt is made to recognize multiply connected characters, both in the form of 8s and in the form of broken characters.

The basic idea is a variant of a well-known gimmick popular among OCR researchers 25 years ago--the method of projection on an axis of the character outline. In this case, it is the left and right extreme profile of the character. Thus, any information within these extreme outlines is neglected. Consider, for example, a 7 with an overhanging tick on the top left; the profile would be the same if there were an extremely thick top bar. Since the authors only deal with numerals, and a small sample at that, this is of no importance; but a larger sample from a larger number of sources, and certainly a larger repertoire of symbols, would cause a breakdown of the method.

In their Abstract, the authors claim that an accuracy of 99 percent was obtained on isolated characters in an “extensive” test experiment. In fact, the experiment concerned only 500 characters obtained from 20 individuals (students in one of the authors’ classes, colleagues?). This means that less than five errors were observed on the test set since connected characters were excluded from this estimate. The confidence level on such a small sample is extremely low.

While there is some discussion of how to segment “simply connected” characters, the discussion of how broken characters (4, 5, and a bizarre variant of 1) are reconnected is rather sketchy. The set of 48 features based on various geometrical (not topological) features of the profiles appears totally ad hoc. For example, if the height-to-width ratio is greater than 2.5, the character is classified as a 1 without further ado.

The training set is not mentioned. It is to be feared that it is the same as the test set, which would of course render totally meaningless the figure of 99 percent recognition on a portion of the latter. In any case, no matter how clever a pattern recognizer may be, it is doubtful that it can ever do better than human beings. As shown by Neisser and Ween [1], using sources at MIT, there is an irreducible minimum of about 3–4 percent of handwritten numerals that humans find strictly ambiguous.

In short, either the work itself is very sloppy or, at least, this paper is. It is up to the reviewer to pick up where the referee has left off.