Evaluation of Transfer Learning for Handwritten Character Classification Using Small Training Samples

In pattern recognition fields, it is worthwhile to develop a pattern recognition system that hears one and knows ten. Recently, classification of printed characters that are the same fonts is almost possible, but classification of handwritten characters is still difficult. On the other hand, there are a large number of writing systems in the world, and there is a need for efficient character classification even with a small sample. Deep learning is one of the most effective approaches for image recognition. Despite this, deep learning causes overtrains easily, particularly when the number of training samples is small. For this reason, deep learning requires a large number of training samples. However, in a practical pattern recognition problem, the number of training samples is usually limited. One method for overcoming this situation is the use of transfer learning, which is pretrained by many samples. In this study, we evaluate the generalization performance of transfer learning for handwritten character classification using a small training sample size. We explore transfer learning using a fine-tuning to fit a small training sample. The experimental results show that transfer learning was more effective for handwritten character classification than convolution neural networks. Transfer learning is expected to be one method that can be used to design a pattern recognition system that works effectively even with a small sample.