Offline Handwriting Recognition of Thai Characters Using Multiple Deep Neural Networks

Handwritten character recognition is a challenge task in many languages. However, languages have their own characteristics. In Thai language, although it is always written from left to right in horizontal direction, the character alignment can be up to four levels in a vertical direction, which is different from several languages such as Chinese, English, Japanese, etc. This work proposed a method for offline handwritten recognition in Thai language based on transfer learning. In this work, we used VGGNet-16 as a pretrained network. The system contains two separate networks for recognizing the characters at base level and the remaining levels which are segmented from a word. Although the character spacing in a word still must be kept enough due to the problems of the segmentation of overlapping characters and the ambiguity of some combinations of adjacent characters in Thai scripts. The proposed method performed quite high accuracies in Thai handwritten recognition.