<![CDATA[Optical Character Recognition (OCR) plays a crucial role in text digitization, yet most existing research predominantly focuses on horizontal English or Latin text, leaving vertical Japanese text largely underexplored. This research addresses that gap by proposing an enhanced Convolutional Recurrent Neural Network (CRNN) model tailored for Japanese vertical text recognition. The novelty of this work lies in integrating residual connections and attention mechanisms into the CRNN, along with a custom dataset that includes both Japanese characters and 30% Latin characters, simulating real-world scenarios such as manga. Experimental results show that the proposed model achieves a Character Error Rate (CER) of 0.363% in only 10 epochs, compared to the baseline CRNN, which reaches 2.084% CER after 100 epochs, demonstrating both faster convergence and improved accuracy. Furthermore, evaluation on manga screenshots from the Manga109 dataset highlights the model’s practical potential, while also revealing new challenges related to irregular vertical spacing and font variability. These findings indicate that the proposed method significantly advances vertical Japanese text recognition and offers a foundation for future improvements in text recognition for manga and similar media.]]>
Copyrights © 2026
