ThaiWritableGAN: Handwriting Generation under Given Information

Abstract

For the local technique challenge, Thai has different symbols’ vertical positions and no space between characters and words. Thai handwriting recognition has been a long time research problem. To join the edge between unsupervised Generative adversarial network (GAN) and Thai handwriting recognition, this paper introduces a novel Thai handwriting generation under given information (named “ThaiWritableGAN”). ThaiWritableGAN is proposed to maps textual information with real handwritten data to generate a new handwritten style (a.k.a. calligraphy). The proposed algorithm consists of generator (G), discriminator (D), and recognizer (R). The synthesized (or generated) handwritten sample is done by G which is proposed to fool D. D is assigned to discriminate an unknown handwritten image that it is real or generated. R is a convolutional neural network (pre-trained by real Thai handwritten images) that is additionally added to recognize the synthesized images (generated by G). For the scientific knowledge discovery, self-attention mechanism (introduced by Google AI) and R totally boost more realistic Thai handwriting generation as well as other languages. The gradient balancing argument should be set to 1. The word error rate (WER) can be relieved by computational reduction in R’s gradient. But the reduction affects a little lower realistic quality of Thai handwriting, measured by Fréchet inception distance (FID). For the beyond following, Thai handwriting generation competition might be opened that the local researchers can submit their handwriting generation algorithms to the calligraphy contest.