Generating a New Dataset for Korean Scene Text Recognition with Augmentation Techniques

초록

Korean text recognition in a natural scene is a challenging task due to the complexity of character shapes and the lack of dataset comparing to English or other languages. In this paper, we present a new dataset with the goal of improving the recognition of Korean natural scene text. Our dataset is generated by data augmentation techniques without losing a reality. The number of augmented images is 3 million and these images are made up of about 30 non-commercial fonts and 511,000 words from a standard Korean language dictionary. This enormous amount of data offers new possibilities for training deeper neural networks. In our extensive experiments, results show that our dataset effectively trains convolutional recurrent neural networks that achieve state-of-the-art performance on the Korea Advanced Institute of Science & Technology (KAIST) scene text database with very few data-acquisition costs.

키워드