The Efficiency of Japanese Character Classification by Attention Mechanism

To improve model performance in the Japanese handwritten character classification task. We designed the most miniaturized but also high-performance CNN-based model. In our initial efforts, we designed a Vanilla CNN architecture and used an attention-based classification head (ML Decoder) to improve accuracy. With 3.6 million parameters, this model achieved an accuracy rate of 99.02% and 97.93% on the test set of the 10-class and 49-class, respectively. After that, we create a network named Residual CNN using the Vanilla CNN model combined with the residual connect of the ResNet network to retain information from the previous layers of the model. After changing the Vanilla CNN backbone to Residual CNN, the model included 3.8M parameters and achieved an accuracy rate of 99.08% and 98.11% on the test set of the 10-class and 49-class, respectively. In addition, the Grad-CAM method is used to improve the explanatory power of how the model makes decisions.