Offline hand-written character recognition by using multinomial logit model

摘要

The Multinomial Logit Model is one of the generalized linear models and the most simplest neural networks for multiple-class pattern recognition. However; few researches are reported those employ this model for image pattern recognition. We employed the model for image pattern (hand-written character) recognition, and estimated its performance. More concretely, we used (1) 36-class data (numbers + English capital letters), (2) 82-class data (numbers + English capital letters + Japanese "katakana" letters) from the ETL6 character-database for experiments. In addition, we applied some techniques to the model to increase recognition rate against unknown data. Experimental results show the effectiveness of combinational utilization of these techniques.