Handwritten Hangul Recognition

Despite significant advancement in Optical Character Recognition (OCR), Handwritten Hangul (Korean) Recognition, abbreviated as HHR, remains largely unsolved due to the similarity found in Hangul handwriting and a much larger number of syllables (∼ 11, 000) to classify compared to English or Latin script languages. The state-of-the-art approach toward HHR on the SERI95a handwritten Korean dataset was based on the AlexNet trained with data augmentation and hybrid learning. While this method has achieved a higher classification performance than the traditional approaches, it is limited in learning feature representations at various levels. In this study, we adopt a variation of Xception which exploits depthwise separable convolutions to enhance the recognition rate of individual syllables. In addition, we propose a progressive training method from Progressive Growing GAN to provide a stabilized training process with reduced training time. Through this approach, the empirical results demonstrate that the proposed approach outperforms the state-of-the-art method and demonstrates its effectiveness on HHR