Thai Handwritten Character Segmentation Based on Deep Learning
Kunnapat Thipparaphonkul, Watchanan Chantapakul, Chayanin Suatap, Karn Patanukhom · 2019
Many computer vision applications rely on segmentation task. To achieve a good result on Handwritten text recognition (HTR), character segmentation is significant in terms of extracting each individual character. In this study, we propose a novel algorithm for tackling offline handwritten character segmentation, particularly for the Thai language. Not only are the characteristics of the Thai language described, but also the problems when performing Thai character segmentation are defined. There are two parts of segmentation: horizontal link segmentation and vertical link segmentation. The chosen type of algorithm is convolutional encoder-decoder network. Our models are based on the renowned encoder-decoder models, U-net and SegNet. The best horizontal link segmentation model achieves up to 0.929 F1-score on the real-world test set. For the vertical link segmentation, the best models of topmost, upper, base, and lower characters attains F1-scores of 0.799, 0.873, 0.932, and 0.820, respectively.