Home > About Us > Technology > Technical Report > Technical Report No.23 2014

Improved Convolutional Neural Network Character Recognition Based on Visual Information Processing Model

The speed of processing digital information is increasing dramatically due to the advances made in IT technology. Conversely, in the field of character recognition, manual confirmation has become the bottleneck, and the recognition rate must be improved. In addition, since the traditional character recognition method that needs manual design to determine the feature vector is time-consuming, using it to change or add such character types as print and handwriting, as well as multiple languages, is difficult. Therefore, a technique called Convolutional Neural Networks (CNN) was invented to automatically determine the feature vector by machine learning. Traditional CNN has the problems of phase invariance and cross-orientation suppression. This report proposes a new method that has achieved a recognition rate of 99.51% for ETL9B from experiments, which is 1.36% higher than the rate achieved by traditional CNN.


  • Masanori Sekino
    Key Technology Laboratory, Research & Technology Group
  • Yutaka Koshi
    Key Technology Laboratory, Research & Technology Group
  • Shunichi Kimura
    Key Technology Laboratory, Research & Technology Group