The basic purpose of the authors is to develop a technique for recognizing “the worst written numerals, as humans can do so easily” rather than implementing an economic “numeral recognition system.”

The paper presents the authors’ results in applying a new method for handwritten-numerals recognition. The method is based on a heuristic technique for extracting the intrinsic structural features of the numerals: the order of interesting points, the numbers of endpoints, forkpoints, crosspoints, breakpoints, and subpieces, the slope-change of each subpiece, the orientation and the length of each subpiece, and the properties of the startpoints and endpoints of each subpiece.

There are five main processing stages for an input image of 128×128 pixels with 256 gray levels: automatic thresholding, thinning, heuristic feature extraction, matching, and identification. The authors implemented their method on a PDP 11/70 system with a Hamamatsu C-1000 camera as input device. The reported overall failure rate is about 3 percent and the average recognition time is about 1s. It is worth mentioning that the method presented in the paper is useful in other recognition applications.