Heuristic approach to handwritten numeral recognition

摘要

An efficient and reliable heuristic method to recognize handwritten numerals is discussed. The numerals are written in a completely random manner such as can be generated and still recognized by humans. The method is designed to read the worst written numerals, where structural differences among numerals are investigated and the extracted heuristic features are: the order of interesting points, numbers of endpoints, forkpoints, crosspoints, breakpoints and subpieces, the slope-change of each subpiece, the orientation and the length of each subpiece, the properties of start- and endpoints of each subpiece. A learning table is built to classify numerals of large variability with very high accuracy.