Improving Recognition of Thai Handwritten Characters with Deep Convolutional Neural Networks

Sarayut Gonwirat and Olarik Surinta

Abstract

For handwritten character recognition, a common problem is that each writer has unique handwriting for each character (e.g. stroke, head, loop, and curl). The similarities of handwritten characters in each language is also a problem. These similarities have led to recognition mistakes. This research compared deep Convolutional Neural Networks (CNNs) which were used for handwriting recognition in the Thai language. CNNs were tested with the THI- C68 dataset. This research also compared two training methods, Train from scratch and Transfer learning, by using VGGNet-19 and Inception-ResNet-v2 architectures. The results showed that VGGNet-19 architecture with transfer learning can reduce learning time. Moreover, it also increased recognition efficiency up to 99.20% when tested with 10-fold cross-validation.

Keywords — Handwritten Character Recognition; Convolutional Neural Network; VGGNet; Inception-ResNet; Transfer Learning.

Read article — https://dl.acm.org/doi/abs/10.1145/3388176.3388181

Thai Dataset

The Thai handwritten character dataset in this research is ALICE- THI dataset, which includes 78. types of Thai characters; consonants, vowels, tones and digits. The dataset contains writing from 150 people, aged 20–23, who were studying in a university. This research used only the THI-68 dataset which eliminated the number. Therefore, the number of characters used in recognition was 68 Characters. The data size was 14,490 characters, including 44 consonants, 17 vowels, 4 tones, and 3 symbols as show in Figure 5.

Surinta et al. used special features siftD take it to learn by SVM algorithm. The accuracy rate was 94.37%. Moreover, Inkeaw et al. used special feature HOGFoDRs with SVM algorithm. The accuracy rate was 98.76%. Due to the similarity of characters, this dataset is challenged for higher effective rate.

Conclusion

This research compares CNN Architectures that are effective in recognizing Thai handwritten characters with a high rate of recognition. The two models are VGGNet-19 and Inception- ResNet-v2 architectures. Both models were evaluated with THI- C68 dataset. In this experiment, the learning method was determined in two types, which are training from scratch and transfer learning. Transfer learning is a way to reduce learning time and increasing the efficiency of recognition. The research has shown that VGGNet-19 architecture with transfer learning has an accuracy rate at 99.20%. In addition, it was higher than Inception- ResNet-v2 architecture. In this regard, VGGNet-19 architecture is a deep learning that has only 19 layers. It has been designed to be stacked together due to make it easier to learn from the network and for increasing the recognition speed.

In future work, researchers will design deep CNN architecture that reduces the number of parameters and reduce learning time. However, the quality must still be equivalent or better performance than with the previous architecture and it will be tested with handwriting characters in other languages such as Bangla, Lanna etc.