Korean Scene Text Recognition Using Semi-Supervised Learning with Character-Level Consistency Regularization
Sung‐Su Kim, Seoung Bum Kim · Journal of Korean Institute of Industrial Engineers · 2023
Scene text recognition is a task that recognizes characters in scene images. Existing studies have been actively conducted based on English but little has been done on Korean. Because Korean does not have enough labeled data and has a large number of characters compared to English, it is more difficult to train Korean scene text recognition model. In addition, most of the previous studies trained the model using synthetic images rather than real images because of insufficient labeled data. However, using synthetic images can reduce generalization performance because of domain gap between real and synthetic images. In this study, we propose a Korean scene text recognition model using semi-supervised learning that overcomes the insufficient labeled data and domain gap. By using text alignment and consistency regularization specialized for Korean scene text recognition, we can obtain better performance than the existing supervised and semi-supervised scene text recognition models for three evaluation datasets. To the best of our knowledge, this is the first study that attempts semi-supervised learning for Korean scene text recognition.