Порівняльний аналіз OCR та LLM підходів для перекладу тексту з зображення сторінки манґи

Abstract

This paper is devoted to the study of using MLLM technology in contextual manga translating field to address the limitation of traditional methods. Two distinct approaches were compared using a custom-made dataset, consisting of original pages and their official English translation. Three models were tested based on translational quality, processing speed and text recognition accuracy. The obtained results showed that MLLM-based approaches can make the process of translation and distribution significantly more efficient. Future work involves developing an application for easy localization