Deep Learning Network-Based End-to-End Scene Text Spotter for Korean Characters
Jeong Won Kim, Seoung Bum Kim · Journal of Korean Institute of Industrial Engineers · 2022
Scene text spotting detects text boxes and recognizes the words from scene images in an end-to-end way. Existing studies proposed scene text spotters for English and showed their promising performance. However, studies for non-English text spotter are rarely conducted because of a lack of fine-quality labeled training datasets. In particular, Korean text spotting is considered to be harder than English because of a large number of characters in Korean. In this study, we propose an end-to-end scene text spotting network specialized in Korean texts that can read more than 2,300 characters. To overcome the lack-of-dataset problem, we propose using a transfer learning method. By pretraining both modules (detector and recognizer) of the network with multi-language datasets and fine-tuning only the recognizer with the Korean dataset, we could construct the robust scene text spotter. We expect that our work can offer useful learning guidelines to future scene text models for non-English language that does not have sufficient training datasets.