STATISTICAL MIXING METHOD FOR AUTOMATIC HANDWRITTEN-CHARACTER RECOGNITION

摘要

PURPOSE: To provide a method and a device for automatically recognizing a handwritten text based upon a suitable handwritten display in mixed decoding for giving consideration to the contribution of one or several feature vector spaces, the Gaussian modeling of respective spaces and all relational prototypes in all the spaces. CONSTITUTION: A feature vector space is selected so as to include local and global description of respective proper points on a pen locus. Window is displayed so as to catch the wide inclination of handwriting and then linear transformation is executed to suitably remove redundancy. A feature vector space obtained as a result is called as handwritten space. A mixing coefficiet for separating sufficient prototype distribution in respective spaces by Gaussian modeling and weighting these distribution by the use of maximum tolerance frame work is trained. Decoding is executed by accumulating the contribution of all relational prototype distribution elements. The method includes post processing using a languge model.