SIMPLE FEATURE EXTRACTION FOR HANDWRITTEN C HARACTER REC 0 GNITION
E. G. Bailey · 1995
This paper deals with a. simple and effective set of features for (hanprinted) cha.ra.cter represeiit,atioii in a.utoinat,ic reading systems. These fea.tures, coinput,ed within regularly pla.ced windows spanning tlie c1ia.ra.ct,er bitinap, consist of a coiiibina.tion of a.vera.ge pixel density aad mea.sures of 1oca.l aligiiiiieiit a.long some directions. Patterns froin different da,taba.ses can be accommodated by choosing a variable window size. These features used in conjuiictioii wit.11 a neural classifier (MLP) yielded a very high a.ccuracy on severa 1 handprint ed char a.c t er d a, tj ab xes, in cl u di ng NIST 's ones. Moreover they are ea.sily iiiipleiiieiit,able in VLSI, with throughputs as high a.s 250,000 cliaract,ers/sec.