A CNN handwritten character recognizer

Abstract
CNNs are used for feature detection in handwritten character recognition. Detected features are fed to a simple classifier network. Performance was tested by using two well‐known ETL data base series: (i) ETL3 consisting of numerals, alphabets and several symbols and (ii) ETL8B2 consisting of Japanese Hirakana characters. the average recognition rate for ETL3 is 94.8%, while that for ETL8B2 is 85.7%. Both series include ‘hard’ characters so distorted that even humans cannot recognize them.

This publication has 7 references indexed in Scilit: